Line data Source code
1 : //! The Page Service listens for client connections and serves their GetPage@LSN
2 : //! requests.
3 :
4 : use std::any::Any;
5 : use std::borrow::Cow;
6 : use std::num::NonZeroUsize;
7 : use std::os::fd::AsRawFd;
8 : use std::pin::Pin;
9 : use std::str::FromStr;
10 : use std::sync::Arc;
11 : use std::task::{Context, Poll};
12 : use std::time::{Duration, Instant, SystemTime};
13 : use std::{io, str};
14 :
15 : use anyhow::{Context as _, bail};
16 : use bytes::{Buf as _, BufMut as _, BytesMut};
17 : use chrono::Utc;
18 : use futures::future::BoxFuture;
19 : use futures::{FutureExt, Stream};
20 : use itertools::Itertools;
21 : use jsonwebtoken::TokenData;
22 : use once_cell::sync::OnceCell;
23 : use pageserver_api::config::{
24 : GetVectoredConcurrentIo, PageServicePipeliningConfig, PageServicePipeliningConfigPipelined,
25 : PageServiceProtocolPipelinedBatchingStrategy, PageServiceProtocolPipelinedExecutionStrategy,
26 : };
27 : use pageserver_api::key::rel_block_to_key;
28 : use pageserver_api::models::{PageTraceEvent, TenantState};
29 : use pageserver_api::pagestream_api::{
30 : self, PagestreamBeMessage, PagestreamDbSizeRequest, PagestreamDbSizeResponse,
31 : PagestreamErrorResponse, PagestreamExistsRequest, PagestreamExistsResponse,
32 : PagestreamFeMessage, PagestreamGetPageRequest, PagestreamGetSlruSegmentRequest,
33 : PagestreamGetSlruSegmentResponse, PagestreamNblocksRequest, PagestreamNblocksResponse,
34 : PagestreamProtocolVersion, PagestreamRequest,
35 : };
36 : use pageserver_api::reltag::SlruKind;
37 : use pageserver_api::shard::TenantShardId;
38 : use pageserver_page_api as page_api;
39 : use pageserver_page_api::proto;
40 : use postgres_backend::{
41 : AuthType, PostgresBackend, PostgresBackendReader, QueryError, is_expected_io_error,
42 : };
43 : use postgres_ffi::BLCKSZ;
44 : use postgres_ffi_types::constants::DEFAULTTABLESPACE_OID;
45 : use pq_proto::framed::ConnectionError;
46 : use pq_proto::{BeMessage, FeMessage, FeStartupPacket, RowDescriptor};
47 : use smallvec::{SmallVec, smallvec};
48 : use strum_macros::IntoStaticStr;
49 : use tokio::io::{AsyncRead, AsyncReadExt as _, AsyncWrite, AsyncWriteExt as _, BufWriter};
50 : use tokio::task::JoinHandle;
51 : use tokio_util::sync::CancellationToken;
52 : use tonic::service::Interceptor as _;
53 : use tonic::transport::server::TcpConnectInfo;
54 : use tracing::*;
55 : use utils::auth::{Claims, Scope, SwappableJwtAuth};
56 : use utils::id::{TenantId, TenantTimelineId, TimelineId};
57 : use utils::logging::log_slow;
58 : use utils::lsn::Lsn;
59 : use utils::shard::ShardIndex;
60 : use utils::simple_rcu::RcuReadGuard;
61 : use utils::sync::gate::{Gate, GateGuard};
62 : use utils::sync::spsc_fold;
63 : use utils::{failpoint_support, span_record};
64 :
65 : use crate::auth::check_permission;
66 : use crate::basebackup::{self, BasebackupError};
67 : use crate::config::PageServerConf;
68 : use crate::context::{
69 : DownloadBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
70 : };
71 : use crate::feature_resolver::FeatureResolver;
72 : use crate::metrics::{
73 : self, COMPUTE_COMMANDS_COUNTERS, ComputeCommandKind, GetPageBatchBreakReason, LIVE_CONNECTIONS,
74 : MISROUTED_PAGESTREAM_REQUESTS, PAGESTREAM_HANDLER_RESULTS_TOTAL, SmgrOpTimer, TimelineMetrics,
75 : };
76 : use crate::pgdatadir_mapping::{LsnRange, Version};
77 : use crate::span::{
78 : debug_assert_current_span_has_tenant_and_timeline_id,
79 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id,
80 : };
81 : use crate::task_mgr::{self, COMPUTE_REQUEST_RUNTIME, TaskKind};
82 : use crate::tenant::mgr::{
83 : GetActiveTenantError, GetTenantError, ShardResolveResult, ShardSelector, TenantManager,
84 : };
85 : use crate::tenant::storage_layer::IoConcurrency;
86 : use crate::tenant::timeline::handle::{Handle, HandleUpgradeError, WeakHandle};
87 : use crate::tenant::timeline::{self, WaitLsnError, WaitLsnTimeout, WaitLsnWaiter};
88 : use crate::tenant::{GetTimelineError, PageReconstructError, Timeline};
89 : use crate::{CancellableTask, PERF_TRACE_TARGET, timed_after_cancellation};
90 :
91 : /// How long we may wait for a [`crate::tenant::mgr::TenantSlot::InProgress`]` and/or a [`crate::tenant::TenantShard`] which
92 : /// is not yet in state [`TenantState::Active`].
93 : ///
94 : /// NB: this is a different value than [`crate::http::routes::ACTIVE_TENANT_TIMEOUT`].
95 : /// HADRON: reduced timeout and we will retry in Cache::get().
96 : const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(5000);
97 :
98 : /// Threshold at which to log slow GetPage requests.
99 : const LOG_SLOW_GETPAGE_THRESHOLD: Duration = Duration::from_secs(30);
100 :
101 : /// The idle time before sending TCP keepalive probes for gRPC connections. The
102 : /// interval and timeout between each probe is configured via sysctl. This
103 : /// allows detecting dead connections sooner.
104 : const GRPC_TCP_KEEPALIVE_TIME: Duration = Duration::from_secs(60);
105 :
106 : /// Whether to enable TCP nodelay for gRPC connections. This disables Nagle's
107 : /// algorithm, which can cause latency spikes for small messages.
108 : const GRPC_TCP_NODELAY: bool = true;
109 :
110 : /// The interval between HTTP2 keepalive pings. This allows shutting down server
111 : /// tasks when clients are unresponsive.
112 : const GRPC_HTTP2_KEEPALIVE_INTERVAL: Duration = Duration::from_secs(30);
113 :
114 : /// The timeout for HTTP2 keepalive pings. Should be <= GRPC_KEEPALIVE_INTERVAL.
115 : const GRPC_HTTP2_KEEPALIVE_TIMEOUT: Duration = Duration::from_secs(20);
116 :
117 : /// Number of concurrent gRPC streams per TCP connection. We expect something
118 : /// like 8 GetPage streams per connections, plus any unary requests.
119 : const GRPC_MAX_CONCURRENT_STREAMS: u32 = 256;
120 :
121 : ///////////////////////////////////////////////////////////////////////////////
122 :
123 : pub struct Listener {
124 : cancel: CancellationToken,
125 : /// Cancel the listener task through `listen_cancel` to shut down the listener
126 : /// and get a handle on the existing connections.
127 : task: JoinHandle<Connections>,
128 : }
129 :
130 : pub struct Connections {
131 : cancel: CancellationToken,
132 : tasks: tokio::task::JoinSet<ConnectionHandlerResult>,
133 : gate: Gate,
134 : }
135 :
136 0 : pub fn spawn(
137 0 : conf: &'static PageServerConf,
138 0 : tenant_manager: Arc<TenantManager>,
139 0 : pg_auth: Option<Arc<SwappableJwtAuth>>,
140 0 : perf_trace_dispatch: Option<Dispatch>,
141 0 : tcp_listener: tokio::net::TcpListener,
142 0 : tls_config: Option<Arc<rustls::ServerConfig>>,
143 0 : feature_resolver: FeatureResolver,
144 0 : ) -> Listener {
145 0 : let cancel = CancellationToken::new();
146 0 : let libpq_ctx = RequestContext::todo_child(
147 0 : TaskKind::LibpqEndpointListener,
148 : // listener task shouldn't need to download anything. (We will
149 : // create a separate sub-contexts for each connection, with their
150 : // own download behavior. This context is used only to listen and
151 : // accept connections.)
152 0 : DownloadBehavior::Error,
153 : );
154 0 : let task = COMPUTE_REQUEST_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
155 0 : "libpq listener",
156 0 : libpq_listener_main(
157 0 : conf,
158 0 : tenant_manager,
159 0 : pg_auth,
160 0 : perf_trace_dispatch,
161 0 : tcp_listener,
162 0 : conf.pg_auth_type,
163 0 : tls_config,
164 0 : conf.page_service_pipelining.clone(),
165 0 : feature_resolver,
166 0 : libpq_ctx,
167 0 : cancel.clone(),
168 0 : )
169 0 : .map(anyhow::Ok),
170 0 : ));
171 :
172 0 : Listener { cancel, task }
173 0 : }
174 :
175 : impl Listener {
176 0 : pub async fn stop_accepting(self) -> Connections {
177 0 : self.cancel.cancel();
178 0 : self.task
179 0 : .await
180 0 : .expect("unreachable: we wrap the listener task in task_mgr::exit_on_panic_or_error")
181 0 : }
182 : }
183 : impl Connections {
184 0 : pub(crate) async fn shutdown(self) {
185 : let Self {
186 0 : cancel,
187 0 : mut tasks,
188 0 : gate,
189 0 : } = self;
190 0 : cancel.cancel();
191 0 : while let Some(res) = tasks.join_next().await {
192 0 : Self::handle_connection_completion(res);
193 0 : }
194 0 : gate.close().await;
195 0 : }
196 :
197 0 : fn handle_connection_completion(res: Result<anyhow::Result<()>, tokio::task::JoinError>) {
198 0 : match res {
199 0 : Ok(Ok(())) => {}
200 0 : Ok(Err(e)) => error!("error in page_service connection task: {:?}", e),
201 0 : Err(e) => error!("page_service connection task panicked: {:?}", e),
202 : }
203 0 : }
204 : }
205 :
206 : ///
207 : /// Main loop of the page service.
208 : ///
209 : /// Listens for connections, and launches a new handler task for each.
210 : ///
211 : /// Returns Ok(()) upon cancellation via `cancel`, returning the set of
212 : /// open connections.
213 : ///
214 : #[allow(clippy::too_many_arguments)]
215 0 : pub async fn libpq_listener_main(
216 0 : conf: &'static PageServerConf,
217 0 : tenant_manager: Arc<TenantManager>,
218 0 : auth: Option<Arc<SwappableJwtAuth>>,
219 0 : perf_trace_dispatch: Option<Dispatch>,
220 0 : listener: tokio::net::TcpListener,
221 0 : auth_type: AuthType,
222 0 : tls_config: Option<Arc<rustls::ServerConfig>>,
223 0 : pipelining_config: PageServicePipeliningConfig,
224 0 : feature_resolver: FeatureResolver,
225 0 : listener_ctx: RequestContext,
226 0 : listener_cancel: CancellationToken,
227 0 : ) -> Connections {
228 0 : let connections_cancel = CancellationToken::new();
229 0 : let connections_gate = Gate::default();
230 0 : let mut connection_handler_tasks = tokio::task::JoinSet::default();
231 :
232 : loop {
233 0 : let gate_guard = match connections_gate.enter() {
234 0 : Ok(guard) => guard,
235 0 : Err(_) => break,
236 : };
237 :
238 0 : let accepted = tokio::select! {
239 : biased;
240 0 : _ = listener_cancel.cancelled() => break,
241 0 : next = connection_handler_tasks.join_next(), if !connection_handler_tasks.is_empty() => {
242 0 : let res = next.expect("we dont poll while empty");
243 0 : Connections::handle_connection_completion(res);
244 0 : continue;
245 : }
246 0 : accepted = listener.accept() => accepted,
247 : };
248 :
249 0 : match accepted {
250 0 : Ok((socket, peer_addr)) => {
251 : // Connection established. Spawn a new task to handle it.
252 0 : debug!("accepted connection from {}", peer_addr);
253 0 : let local_auth = auth.clone();
254 0 : let connection_ctx = RequestContextBuilder::from(&listener_ctx)
255 0 : .task_kind(TaskKind::PageRequestHandler)
256 0 : .download_behavior(DownloadBehavior::Download)
257 0 : .perf_span_dispatch(perf_trace_dispatch.clone())
258 0 : .detached_child();
259 :
260 0 : connection_handler_tasks.spawn(page_service_conn_main(
261 0 : conf,
262 0 : tenant_manager.clone(),
263 0 : local_auth,
264 0 : socket,
265 0 : auth_type,
266 0 : tls_config.clone(),
267 0 : pipelining_config.clone(),
268 0 : feature_resolver.clone(),
269 0 : connection_ctx,
270 0 : connections_cancel.child_token(),
271 0 : gate_guard,
272 : ));
273 : }
274 0 : Err(err) => {
275 : // accept() failed. Log the error, and loop back to retry on next connection.
276 0 : error!("accept() failed: {:?}", err);
277 : }
278 : }
279 : }
280 :
281 0 : debug!("page_service listener loop terminated");
282 :
283 0 : Connections {
284 0 : cancel: connections_cancel,
285 0 : tasks: connection_handler_tasks,
286 0 : gate: connections_gate,
287 0 : }
288 0 : }
289 :
290 : type ConnectionHandlerResult = anyhow::Result<()>;
291 :
292 : /// Perf root spans start at the per-request level, after shard routing.
293 : /// This struct carries connection-level information to the root perf span definition.
294 : #[derive(Clone, Default)]
295 : struct ConnectionPerfSpanFields {
296 : peer_addr: String,
297 : application_name: Option<String>,
298 : compute_mode: Option<String>,
299 : }
300 :
301 : #[instrument(skip_all, fields(peer_addr, application_name, compute_mode))]
302 : #[allow(clippy::too_many_arguments)]
303 : async fn page_service_conn_main(
304 : conf: &'static PageServerConf,
305 : tenant_manager: Arc<TenantManager>,
306 : auth: Option<Arc<SwappableJwtAuth>>,
307 : socket: tokio::net::TcpStream,
308 : auth_type: AuthType,
309 : tls_config: Option<Arc<rustls::ServerConfig>>,
310 : pipelining_config: PageServicePipeliningConfig,
311 : feature_resolver: FeatureResolver,
312 : connection_ctx: RequestContext,
313 : cancel: CancellationToken,
314 : gate_guard: GateGuard,
315 : ) -> ConnectionHandlerResult {
316 : let _guard = LIVE_CONNECTIONS
317 : .with_label_values(&["page_service"])
318 : .guard();
319 :
320 : socket
321 : .set_nodelay(true)
322 : .context("could not set TCP_NODELAY")?;
323 :
324 : let socket_fd = socket.as_raw_fd();
325 :
326 : let peer_addr = socket.peer_addr().context("get peer address")?;
327 :
328 : let perf_span_fields = ConnectionPerfSpanFields {
329 : peer_addr: peer_addr.to_string(),
330 : application_name: None, // filled in later
331 : compute_mode: None, // filled in later
332 : };
333 : tracing::Span::current().record("peer_addr", field::display(peer_addr));
334 :
335 : // setup read timeout of 10 minutes. the timeout is rather arbitrary for requirements:
336 : // - long enough for most valid compute connections
337 : // - less than infinite to stop us from "leaking" connections to long-gone computes
338 : //
339 : // no write timeout is used, because the kernel is assumed to error writes after some time.
340 : let mut socket = tokio_io_timeout::TimeoutReader::new(socket);
341 :
342 : let default_timeout_ms = 10 * 60 * 1000; // 10 minutes by default
343 0 : let socket_timeout_ms = (|| {
344 0 : fail::fail_point!("simulated-bad-compute-connection", |avg_timeout_ms| {
345 : // Exponential distribution for simulating
346 : // poor network conditions, expect about avg_timeout_ms to be around 15
347 : // in tests
348 0 : if let Some(avg_timeout_ms) = avg_timeout_ms {
349 0 : let avg = avg_timeout_ms.parse::<i64>().unwrap() as f32;
350 0 : let u = rand::random::<f32>();
351 0 : ((1.0 - u).ln() / (-avg)) as u64
352 : } else {
353 0 : default_timeout_ms
354 : }
355 0 : });
356 0 : default_timeout_ms
357 : })();
358 :
359 : // A timeout here does not mean the client died, it can happen if it's just idle for
360 : // a while: we will tear down this PageServerHandler and instantiate a new one if/when
361 : // they reconnect.
362 : socket.set_timeout(Some(std::time::Duration::from_millis(socket_timeout_ms)));
363 : let socket = Box::pin(socket);
364 :
365 : fail::fail_point!("ps::connection-start::pre-login");
366 :
367 : // XXX: pgbackend.run() should take the connection_ctx,
368 : // and create a child per-query context when it invokes process_query.
369 : // But it's in a shared crate, so, we store connection_ctx inside PageServerHandler
370 : // and create the per-query context in process_query ourselves.
371 : let mut conn_handler = PageServerHandler::new(
372 : tenant_manager,
373 : auth,
374 : pipelining_config,
375 : conf.get_vectored_concurrent_io,
376 : perf_span_fields,
377 : connection_ctx,
378 : cancel.clone(),
379 : feature_resolver.clone(),
380 : gate_guard,
381 : );
382 : let pgbackend =
383 : PostgresBackend::new_from_io(socket_fd, socket, peer_addr, auth_type, tls_config)?;
384 :
385 : match pgbackend.run(&mut conn_handler, &cancel).await {
386 : Ok(()) => {
387 : // we've been requested to shut down
388 : Ok(())
389 : }
390 : Err(QueryError::Disconnected(ConnectionError::Io(io_error))) => {
391 : if is_expected_io_error(&io_error) {
392 : info!("Postgres client disconnected ({io_error})");
393 : Ok(())
394 : } else {
395 : let tenant_id = conn_handler.timeline_handles.as_ref().unwrap().tenant_id();
396 : Err(io_error).context(format!(
397 : "Postgres connection error for tenant_id={tenant_id:?} client at peer_addr={peer_addr}"
398 : ))
399 : }
400 : }
401 : other => {
402 : let tenant_id = conn_handler.timeline_handles.as_ref().unwrap().tenant_id();
403 : other.context(format!(
404 : "Postgres query error for tenant_id={tenant_id:?} client peer_addr={peer_addr}"
405 : ))
406 : }
407 : }
408 : }
409 :
410 : /// Page service connection handler.
411 : struct PageServerHandler {
412 : auth: Option<Arc<SwappableJwtAuth>>,
413 : claims: Option<Claims>,
414 :
415 : /// The context created for the lifetime of the connection
416 : /// services by this PageServerHandler.
417 : /// For each query received over the connection,
418 : /// `process_query` creates a child context from this one.
419 : connection_ctx: RequestContext,
420 :
421 : perf_span_fields: ConnectionPerfSpanFields,
422 :
423 : cancel: CancellationToken,
424 :
425 : /// None only while pagestream protocol is being processed.
426 : timeline_handles: Option<TimelineHandles>,
427 :
428 : pipelining_config: PageServicePipeliningConfig,
429 : get_vectored_concurrent_io: GetVectoredConcurrentIo,
430 :
431 : feature_resolver: FeatureResolver,
432 :
433 : gate_guard: GateGuard,
434 : }
435 :
436 : struct TimelineHandles {
437 : wrapper: TenantManagerWrapper,
438 : /// Note on size: the typical size of this map is 1. The largest size we expect
439 : /// to see is the number of shards divided by the number of pageservers (typically < 2),
440 : /// or the ratio used when splitting shards (i.e. how many children created from one)
441 : /// parent shard, where a "large" number might be ~8.
442 : handles: timeline::handle::Cache<TenantManagerTypes>,
443 : }
444 :
445 : impl TimelineHandles {
446 0 : fn new(tenant_manager: Arc<TenantManager>) -> Self {
447 0 : Self {
448 0 : wrapper: TenantManagerWrapper {
449 0 : tenant_manager,
450 0 : tenant_id: OnceCell::new(),
451 0 : },
452 0 : handles: Default::default(),
453 0 : }
454 0 : }
455 0 : async fn get(
456 0 : &mut self,
457 0 : tenant_id: TenantId,
458 0 : timeline_id: TimelineId,
459 0 : shard_selector: ShardSelector,
460 0 : ) -> Result<Handle<TenantManagerTypes>, GetActiveTimelineError> {
461 0 : if *self.wrapper.tenant_id.get_or_init(|| tenant_id) != tenant_id {
462 0 : return Err(GetActiveTimelineError::Tenant(
463 0 : GetActiveTenantError::SwitchedTenant,
464 0 : ));
465 0 : }
466 0 : self.handles
467 0 : .get(timeline_id, shard_selector, &self.wrapper)
468 0 : .await
469 0 : }
470 :
471 0 : fn tenant_id(&self) -> Option<TenantId> {
472 0 : self.wrapper.tenant_id.get().copied()
473 0 : }
474 : }
475 :
476 : pub(crate) struct TenantManagerWrapper {
477 : tenant_manager: Arc<TenantManager>,
478 : // We do not support switching tenant_id on a connection at this point.
479 : // We can can add support for this later if needed without changing
480 : // the protocol.
481 : tenant_id: once_cell::sync::OnceCell<TenantId>,
482 : }
483 :
484 : pub(crate) struct TenantManagerTypes;
485 :
486 : impl timeline::handle::Types for TenantManagerTypes {
487 : type TenantManager = TenantManagerWrapper;
488 : type Timeline = TenantManagerCacheItem;
489 : }
490 :
491 : pub(crate) struct TenantManagerCacheItem {
492 : pub(crate) timeline: Arc<Timeline>,
493 : // allow() for cheap propagation through RequestContext inside a task
494 : #[allow(clippy::redundant_allocation)]
495 : pub(crate) metrics: Arc<Arc<TimelineMetrics>>,
496 : #[allow(dead_code)] // we store it to keep the gate open
497 : pub(crate) gate_guard: GateGuard,
498 : }
499 :
500 : impl std::ops::Deref for TenantManagerCacheItem {
501 : type Target = Arc<Timeline>;
502 0 : fn deref(&self) -> &Self::Target {
503 0 : &self.timeline
504 0 : }
505 : }
506 :
507 : impl timeline::handle::Timeline<TenantManagerTypes> for TenantManagerCacheItem {
508 0 : fn shard_timeline_id(&self) -> timeline::handle::ShardTimelineId {
509 0 : Timeline::shard_timeline_id(&self.timeline)
510 0 : }
511 :
512 0 : fn per_timeline_state(&self) -> &timeline::handle::PerTimelineState<TenantManagerTypes> {
513 0 : &self.timeline.handles
514 0 : }
515 :
516 0 : fn get_shard_identity(&self) -> &pageserver_api::shard::ShardIdentity {
517 0 : Timeline::get_shard_identity(&self.timeline)
518 0 : }
519 : }
520 :
521 : impl timeline::handle::TenantManager<TenantManagerTypes> for TenantManagerWrapper {
522 0 : async fn resolve(
523 0 : &self,
524 0 : timeline_id: TimelineId,
525 0 : shard_selector: ShardSelector,
526 0 : ) -> Result<TenantManagerCacheItem, GetActiveTimelineError> {
527 0 : let tenant_id = self.tenant_id.get().expect("we set this in get()");
528 0 : let timeout = ACTIVE_TENANT_TIMEOUT;
529 0 : let wait_start = Instant::now();
530 0 : let deadline = wait_start + timeout;
531 0 : let tenant_shard = loop {
532 0 : let resolved = self
533 0 : .tenant_manager
534 0 : .resolve_attached_shard(tenant_id, shard_selector);
535 0 : match resolved {
536 0 : ShardResolveResult::Found(tenant_shard) => break tenant_shard,
537 : ShardResolveResult::NotFound => {
538 0 : MISROUTED_PAGESTREAM_REQUESTS.inc();
539 0 : return Err(GetActiveTimelineError::Tenant(
540 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(*tenant_id)),
541 0 : ));
542 : }
543 0 : ShardResolveResult::InProgress(barrier) => {
544 : // We can't authoritatively answer right now: wait for InProgress state
545 : // to end, then try again
546 0 : tokio::select! {
547 0 : _ = barrier.wait() => {
548 0 : // The barrier completed: proceed around the loop to try looking up again
549 0 : },
550 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
551 0 : return Err(GetActiveTimelineError::Tenant(GetActiveTenantError::WaitForActiveTimeout {
552 0 : latest_state: None,
553 0 : wait_time: timeout,
554 0 : }));
555 : }
556 : }
557 : }
558 : };
559 : };
560 :
561 0 : tracing::debug!("Waiting for tenant to enter active state...");
562 0 : tenant_shard
563 0 : .wait_to_become_active(deadline.duration_since(Instant::now()))
564 0 : .await
565 0 : .map_err(GetActiveTimelineError::Tenant)?;
566 :
567 0 : let timeline = tenant_shard
568 0 : .get_timeline(timeline_id, true)
569 0 : .map_err(GetActiveTimelineError::Timeline)?;
570 :
571 0 : let gate_guard = match timeline.gate.enter() {
572 0 : Ok(guard) => guard,
573 : Err(_) => {
574 0 : return Err(GetActiveTimelineError::Timeline(
575 0 : GetTimelineError::ShuttingDown,
576 0 : ));
577 : }
578 : };
579 :
580 0 : let metrics = Arc::new(Arc::clone(&timeline.metrics));
581 :
582 0 : Ok(TenantManagerCacheItem {
583 0 : timeline,
584 0 : metrics,
585 0 : gate_guard,
586 0 : })
587 0 : }
588 : }
589 :
590 : /// Whether to hold the applied GC cutoff guard when processing GetPage requests.
591 : /// This is determined once at the start of pagestream subprotocol handling based on
592 : /// feature flags, configuration, and test conditions.
593 : #[derive(Debug, Clone, Copy)]
594 : enum HoldAppliedGcCutoffGuard {
595 : Yes,
596 : No,
597 : }
598 :
599 : #[derive(thiserror::Error, Debug)]
600 : enum PageStreamError {
601 : /// We encountered an error that should prompt the client to reconnect:
602 : /// in practice this means we drop the connection without sending a response.
603 : #[error("Reconnect required: {0}")]
604 : Reconnect(Cow<'static, str>),
605 :
606 : /// We were instructed to shutdown while processing the query
607 : #[error("Shutting down")]
608 : Shutdown,
609 :
610 : /// Something went wrong reading a page: this likely indicates a pageserver bug
611 : #[error("Read error")]
612 : Read(#[source] PageReconstructError),
613 :
614 : /// Ran out of time waiting for an LSN
615 : #[error("LSN timeout: {0}")]
616 : LsnTimeout(WaitLsnError),
617 :
618 : /// The entity required to serve the request (tenant or timeline) is not found,
619 : /// or is not found in a suitable state to serve a request.
620 : #[error("Not found: {0}")]
621 : NotFound(Cow<'static, str>),
622 :
623 : /// Request asked for something that doesn't make sense, like an invalid LSN
624 : #[error("Bad request: {0}")]
625 : BadRequest(Cow<'static, str>),
626 : }
627 :
628 : impl From<PageStreamError> for tonic::Status {
629 0 : fn from(err: PageStreamError) -> Self {
630 : use tonic::Code;
631 0 : let message = err.to_string();
632 0 : let code = match err {
633 0 : PageStreamError::Reconnect(_) => Code::Unavailable,
634 0 : PageStreamError::Shutdown => Code::Unavailable,
635 0 : PageStreamError::Read(err) => match err {
636 0 : PageReconstructError::Cancelled => Code::Unavailable,
637 0 : PageReconstructError::MissingKey(_) => Code::NotFound,
638 0 : PageReconstructError::AncestorLsnTimeout(err) => tonic::Status::from(err).code(),
639 0 : PageReconstructError::Other(_) => Code::Internal,
640 0 : PageReconstructError::WalRedo(_) => Code::Internal,
641 : },
642 0 : PageStreamError::LsnTimeout(err) => tonic::Status::from(err).code(),
643 0 : PageStreamError::NotFound(_) => Code::NotFound,
644 0 : PageStreamError::BadRequest(_) => Code::InvalidArgument,
645 : };
646 0 : tonic::Status::new(code, message)
647 0 : }
648 : }
649 :
650 : impl From<PageReconstructError> for PageStreamError {
651 0 : fn from(value: PageReconstructError) -> Self {
652 0 : match value {
653 0 : PageReconstructError::Cancelled => Self::Shutdown,
654 0 : e => Self::Read(e),
655 : }
656 0 : }
657 : }
658 :
659 : impl From<GetActiveTimelineError> for PageStreamError {
660 0 : fn from(value: GetActiveTimelineError) -> Self {
661 0 : match value {
662 : GetActiveTimelineError::Tenant(GetActiveTenantError::Cancelled)
663 : | GetActiveTimelineError::Tenant(GetActiveTenantError::WillNotBecomeActive(
664 : TenantState::Stopping { .. },
665 : ))
666 0 : | GetActiveTimelineError::Timeline(GetTimelineError::ShuttingDown) => Self::Shutdown,
667 0 : GetActiveTimelineError::Tenant(e) => Self::NotFound(format!("{e}").into()),
668 0 : GetActiveTimelineError::Timeline(e) => Self::NotFound(format!("{e}").into()),
669 : }
670 0 : }
671 : }
672 :
673 : impl From<WaitLsnError> for PageStreamError {
674 0 : fn from(value: WaitLsnError) -> Self {
675 0 : match value {
676 0 : e @ WaitLsnError::Timeout(_) => Self::LsnTimeout(e),
677 0 : WaitLsnError::Shutdown => Self::Shutdown,
678 0 : e @ WaitLsnError::BadState { .. } => Self::Reconnect(format!("{e}").into()),
679 : }
680 0 : }
681 : }
682 :
683 : impl From<WaitLsnError> for QueryError {
684 0 : fn from(value: WaitLsnError) -> Self {
685 0 : match value {
686 0 : e @ WaitLsnError::Timeout(_) => Self::Other(anyhow::Error::new(e)),
687 0 : WaitLsnError::Shutdown => Self::Shutdown,
688 0 : WaitLsnError::BadState { .. } => Self::Reconnect,
689 : }
690 0 : }
691 : }
692 :
693 : #[derive(thiserror::Error, Debug)]
694 : struct BatchedPageStreamError {
695 : req: PagestreamRequest,
696 : err: PageStreamError,
697 : }
698 :
699 : impl std::fmt::Display for BatchedPageStreamError {
700 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
701 0 : self.err.fmt(f)
702 0 : }
703 : }
704 :
705 : struct BatchedGetPageRequest {
706 : req: PagestreamGetPageRequest,
707 : timer: SmgrOpTimer,
708 : lsn_range: LsnRange,
709 : ctx: RequestContext,
710 : // If the request is perf enabled, this contains a context
711 : // with a perf span tracking the time spent waiting for the executor.
712 : batch_wait_ctx: Option<RequestContext>,
713 : }
714 :
715 : #[cfg(feature = "testing")]
716 : struct BatchedTestRequest {
717 : req: pagestream_api::PagestreamTestRequest,
718 : timer: SmgrOpTimer,
719 : }
720 :
721 : /// NB: we only hold [`timeline::handle::WeakHandle`] inside this enum,
722 : /// so that we don't keep the [`Timeline::gate`] open while the batch
723 : /// is being built up inside the [`spsc_fold`] (pagestream pipelining).
724 : #[derive(IntoStaticStr)]
725 : #[allow(clippy::large_enum_variant)]
726 : enum BatchedFeMessage {
727 : Exists {
728 : span: Span,
729 : timer: SmgrOpTimer,
730 : shard: WeakHandle<TenantManagerTypes>,
731 : req: PagestreamExistsRequest,
732 : },
733 : Nblocks {
734 : span: Span,
735 : timer: SmgrOpTimer,
736 : shard: WeakHandle<TenantManagerTypes>,
737 : req: PagestreamNblocksRequest,
738 : },
739 : GetPage {
740 : span: Span,
741 : shard: WeakHandle<TenantManagerTypes>,
742 : applied_gc_cutoff_guard: Option<RcuReadGuard<Lsn>>,
743 : pages: SmallVec<[BatchedGetPageRequest; 1]>,
744 : batch_break_reason: GetPageBatchBreakReason,
745 : },
746 : DbSize {
747 : span: Span,
748 : timer: SmgrOpTimer,
749 : shard: WeakHandle<TenantManagerTypes>,
750 : req: PagestreamDbSizeRequest,
751 : },
752 : GetSlruSegment {
753 : span: Span,
754 : timer: SmgrOpTimer,
755 : shard: WeakHandle<TenantManagerTypes>,
756 : req: PagestreamGetSlruSegmentRequest,
757 : },
758 : #[cfg(feature = "testing")]
759 : Test {
760 : span: Span,
761 : shard: WeakHandle<TenantManagerTypes>,
762 : requests: Vec<BatchedTestRequest>,
763 : },
764 : RespondError {
765 : span: Span,
766 : error: BatchedPageStreamError,
767 : },
768 : }
769 :
770 : impl BatchedFeMessage {
771 0 : fn as_static_str(&self) -> &'static str {
772 0 : self.into()
773 0 : }
774 :
775 0 : fn observe_execution_start(&mut self, at: Instant) {
776 0 : match self {
777 0 : BatchedFeMessage::Exists { timer, .. }
778 0 : | BatchedFeMessage::Nblocks { timer, .. }
779 0 : | BatchedFeMessage::DbSize { timer, .. }
780 0 : | BatchedFeMessage::GetSlruSegment { timer, .. } => {
781 0 : timer.observe_execution_start(at);
782 0 : }
783 0 : BatchedFeMessage::GetPage { pages, .. } => {
784 0 : for page in pages {
785 0 : page.timer.observe_execution_start(at);
786 0 : }
787 : }
788 : #[cfg(feature = "testing")]
789 0 : BatchedFeMessage::Test { requests, .. } => {
790 0 : for req in requests {
791 0 : req.timer.observe_execution_start(at);
792 0 : }
793 : }
794 0 : BatchedFeMessage::RespondError { .. } => {}
795 : }
796 0 : }
797 :
798 0 : fn should_break_batch(
799 0 : &self,
800 0 : other: &BatchedFeMessage,
801 0 : max_batch_size: NonZeroUsize,
802 0 : batching_strategy: PageServiceProtocolPipelinedBatchingStrategy,
803 0 : ) -> Option<GetPageBatchBreakReason> {
804 0 : match (self, other) {
805 : (
806 : BatchedFeMessage::GetPage {
807 0 : shard: accum_shard,
808 0 : pages: accum_pages,
809 : ..
810 : },
811 : BatchedFeMessage::GetPage {
812 0 : shard: this_shard,
813 0 : pages: this_pages,
814 : ..
815 : },
816 : ) => {
817 0 : assert_eq!(this_pages.len(), 1);
818 0 : if accum_pages.len() >= max_batch_size.get() {
819 0 : trace!(%max_batch_size, "stopping batching because of batch size");
820 0 : assert_eq!(accum_pages.len(), max_batch_size.get());
821 :
822 0 : return Some(GetPageBatchBreakReason::BatchFull);
823 0 : }
824 0 : if !accum_shard.is_same_handle_as(this_shard) {
825 0 : trace!("stopping batching because timeline object mismatch");
826 : // TODO: we _could_ batch & execute each shard seperately (and in parallel).
827 : // But the current logic for keeping responses in order does not support that.
828 :
829 0 : return Some(GetPageBatchBreakReason::NonUniformTimeline);
830 0 : }
831 :
832 0 : match batching_strategy {
833 : PageServiceProtocolPipelinedBatchingStrategy::UniformLsn => {
834 0 : if let Some(last_in_batch) = accum_pages.last() {
835 0 : if last_in_batch.lsn_range.effective_lsn
836 0 : != this_pages[0].lsn_range.effective_lsn
837 : {
838 0 : trace!(
839 : accum_lsn = %last_in_batch.lsn_range.effective_lsn,
840 0 : this_lsn = %this_pages[0].lsn_range.effective_lsn,
841 0 : "stopping batching because LSN changed"
842 : );
843 :
844 0 : return Some(GetPageBatchBreakReason::NonUniformLsn);
845 0 : }
846 0 : }
847 : }
848 : PageServiceProtocolPipelinedBatchingStrategy::ScatteredLsn => {
849 : // The read path doesn't curently support serving the same page at different LSNs.
850 : // While technically possible, it's uncertain if the complexity is worth it.
851 : // Break the batch if such a case is encountered.
852 0 : let same_page_different_lsn = accum_pages.iter().any(|batched| {
853 0 : batched.req.rel == this_pages[0].req.rel
854 0 : && batched.req.blkno == this_pages[0].req.blkno
855 0 : && batched.lsn_range.effective_lsn
856 0 : != this_pages[0].lsn_range.effective_lsn
857 0 : });
858 :
859 0 : if same_page_different_lsn {
860 0 : trace!(
861 0 : rel=%this_pages[0].req.rel,
862 0 : blkno=%this_pages[0].req.blkno,
863 0 : lsn=%this_pages[0].lsn_range.effective_lsn,
864 0 : "stopping batching because same page was requested at different LSNs"
865 : );
866 :
867 0 : return Some(GetPageBatchBreakReason::SamePageAtDifferentLsn);
868 0 : }
869 : }
870 : }
871 :
872 0 : None
873 : }
874 : #[cfg(feature = "testing")]
875 : (
876 : BatchedFeMessage::Test {
877 0 : shard: accum_shard,
878 0 : requests: accum_requests,
879 : ..
880 : },
881 : BatchedFeMessage::Test {
882 0 : shard: this_shard,
883 0 : requests: this_requests,
884 : ..
885 : },
886 : ) => {
887 0 : assert!(this_requests.len() == 1);
888 0 : if accum_requests.len() >= max_batch_size.get() {
889 0 : trace!(%max_batch_size, "stopping batching because of batch size");
890 0 : assert_eq!(accum_requests.len(), max_batch_size.get());
891 0 : return Some(GetPageBatchBreakReason::BatchFull);
892 0 : }
893 0 : if !accum_shard.is_same_handle_as(this_shard) {
894 0 : trace!("stopping batching because timeline object mismatch");
895 : // TODO: we _could_ batch & execute each shard seperately (and in parallel).
896 : // But the current logic for keeping responses in order does not support that.
897 0 : return Some(GetPageBatchBreakReason::NonUniformTimeline);
898 0 : }
899 0 : let this_batch_key = this_requests[0].req.batch_key;
900 0 : let accum_batch_key = accum_requests[0].req.batch_key;
901 0 : if this_requests[0].req.batch_key != accum_requests[0].req.batch_key {
902 0 : trace!(%accum_batch_key, %this_batch_key, "stopping batching because batch key changed");
903 0 : return Some(GetPageBatchBreakReason::NonUniformKey);
904 0 : }
905 0 : None
906 : }
907 0 : (_, _) => Some(GetPageBatchBreakReason::NonBatchableRequest),
908 : }
909 0 : }
910 : }
911 :
912 : impl PageServerHandler {
913 : #[allow(clippy::too_many_arguments)]
914 0 : pub fn new(
915 0 : tenant_manager: Arc<TenantManager>,
916 0 : auth: Option<Arc<SwappableJwtAuth>>,
917 0 : pipelining_config: PageServicePipeliningConfig,
918 0 : get_vectored_concurrent_io: GetVectoredConcurrentIo,
919 0 : perf_span_fields: ConnectionPerfSpanFields,
920 0 : connection_ctx: RequestContext,
921 0 : cancel: CancellationToken,
922 0 : feature_resolver: FeatureResolver,
923 0 : gate_guard: GateGuard,
924 0 : ) -> Self {
925 0 : PageServerHandler {
926 0 : auth,
927 0 : claims: None,
928 0 : connection_ctx,
929 0 : perf_span_fields,
930 0 : timeline_handles: Some(TimelineHandles::new(tenant_manager)),
931 0 : cancel,
932 0 : pipelining_config,
933 0 : get_vectored_concurrent_io,
934 0 : feature_resolver,
935 0 : gate_guard,
936 0 : }
937 0 : }
938 :
939 : /// This function always respects cancellation of any timeline in `[Self::shard_timelines]`. Pass in
940 : /// a cancellation token at the next scope up (such as a tenant cancellation token) to ensure we respect
941 : /// cancellation if there aren't any timelines in the cache.
942 : ///
943 : /// If calling from a function that doesn't use the `[Self::shard_timelines]` cache, then pass in the
944 : /// timeline cancellation token.
945 0 : async fn flush_cancellable<IO>(
946 0 : &self,
947 0 : pgb: &mut PostgresBackend<IO>,
948 0 : cancel: &CancellationToken,
949 0 : ) -> Result<(), QueryError>
950 0 : where
951 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
952 0 : {
953 0 : tokio::select!(
954 0 : flush_r = pgb.flush() => {
955 0 : Ok(flush_r?)
956 : },
957 0 : _ = cancel.cancelled() => {
958 0 : Err(QueryError::Shutdown)
959 : }
960 : )
961 0 : }
962 :
963 : #[allow(clippy::too_many_arguments)]
964 0 : async fn pagestream_read_message<IO>(
965 0 : pgb: &mut PostgresBackendReader<IO>,
966 0 : tenant_id: TenantId,
967 0 : timeline_id: TimelineId,
968 0 : timeline_handles: &mut TimelineHandles,
969 0 : conn_perf_span_fields: &ConnectionPerfSpanFields,
970 0 : cancel: &CancellationToken,
971 0 : ctx: &RequestContext,
972 0 : protocol_version: PagestreamProtocolVersion,
973 0 : parent_span: Span,
974 0 : hold_gc_cutoff_guard: HoldAppliedGcCutoffGuard,
975 0 : ) -> Result<Option<BatchedFeMessage>, QueryError>
976 0 : where
977 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
978 0 : {
979 0 : let msg = tokio::select! {
980 : biased;
981 0 : _ = cancel.cancelled() => {
982 0 : return Err(QueryError::Shutdown)
983 : }
984 0 : msg = pgb.read_message() => { msg }
985 : };
986 :
987 0 : let received_at = Instant::now();
988 :
989 0 : let copy_data_bytes = match msg? {
990 0 : Some(FeMessage::CopyData(bytes)) => bytes,
991 : Some(FeMessage::Terminate) => {
992 0 : return Ok(None);
993 : }
994 0 : Some(m) => {
995 0 : return Err(QueryError::Other(anyhow::anyhow!(
996 0 : "unexpected message: {m:?} during COPY"
997 0 : )));
998 : }
999 : None => {
1000 0 : return Ok(None);
1001 : } // client disconnected
1002 : };
1003 0 : trace!("query: {copy_data_bytes:?}");
1004 :
1005 0 : fail::fail_point!("ps::handle-pagerequest-message");
1006 :
1007 : // parse request
1008 0 : let neon_fe_msg =
1009 0 : PagestreamFeMessage::parse(&mut copy_data_bytes.reader(), protocol_version)?;
1010 :
1011 0 : let batched_msg = match neon_fe_msg {
1012 0 : PagestreamFeMessage::Exists(req) => {
1013 0 : let shard = timeline_handles
1014 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
1015 0 : .await?;
1016 0 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
1017 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());
1018 0 : let timer = Self::record_op_start_and_throttle(
1019 0 : &shard,
1020 0 : metrics::SmgrQueryType::GetRelExists,
1021 0 : received_at,
1022 0 : )
1023 0 : .await?;
1024 0 : BatchedFeMessage::Exists {
1025 0 : span,
1026 0 : timer,
1027 0 : shard: shard.downgrade(),
1028 0 : req,
1029 0 : }
1030 : }
1031 0 : PagestreamFeMessage::Nblocks(req) => {
1032 0 : let shard = timeline_handles
1033 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
1034 0 : .await?;
1035 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());
1036 0 : let timer = Self::record_op_start_and_throttle(
1037 0 : &shard,
1038 0 : metrics::SmgrQueryType::GetRelSize,
1039 0 : received_at,
1040 0 : )
1041 0 : .await?;
1042 0 : BatchedFeMessage::Nblocks {
1043 0 : span,
1044 0 : timer,
1045 0 : shard: shard.downgrade(),
1046 0 : req,
1047 0 : }
1048 : }
1049 0 : PagestreamFeMessage::DbSize(req) => {
1050 0 : let shard = timeline_handles
1051 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
1052 0 : .await?;
1053 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());
1054 0 : let timer = Self::record_op_start_and_throttle(
1055 0 : &shard,
1056 0 : metrics::SmgrQueryType::GetDbSize,
1057 0 : received_at,
1058 0 : )
1059 0 : .await?;
1060 0 : BatchedFeMessage::DbSize {
1061 0 : span,
1062 0 : timer,
1063 0 : shard: shard.downgrade(),
1064 0 : req,
1065 0 : }
1066 : }
1067 0 : PagestreamFeMessage::GetSlruSegment(req) => {
1068 0 : let shard = timeline_handles
1069 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
1070 0 : .await?;
1071 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());
1072 0 : let timer = Self::record_op_start_and_throttle(
1073 0 : &shard,
1074 0 : metrics::SmgrQueryType::GetSlruSegment,
1075 0 : received_at,
1076 0 : )
1077 0 : .await?;
1078 0 : BatchedFeMessage::GetSlruSegment {
1079 0 : span,
1080 0 : timer,
1081 0 : shard: shard.downgrade(),
1082 0 : req,
1083 0 : }
1084 : }
1085 0 : PagestreamFeMessage::GetPage(req) => {
1086 : // avoid a somewhat costly Span::record() by constructing the entire span in one go.
1087 : macro_rules! mkspan {
1088 : (before shard routing) => {{
1089 : tracing::info_span!(
1090 : parent: &parent_span,
1091 : "handle_get_page_request",
1092 : request_id = %req.hdr.reqid,
1093 : rel = %req.rel,
1094 : blkno = %req.blkno,
1095 : req_lsn = %req.hdr.request_lsn,
1096 : not_modified_since_lsn = %req.hdr.not_modified_since,
1097 : )
1098 : }};
1099 : ($shard_id:expr) => {{
1100 : tracing::info_span!(
1101 : parent: &parent_span,
1102 : "handle_get_page_request",
1103 : request_id = %req.hdr.reqid,
1104 : rel = %req.rel,
1105 : blkno = %req.blkno,
1106 : req_lsn = %req.hdr.request_lsn,
1107 : not_modified_since_lsn = %req.hdr.not_modified_since,
1108 : shard_id = %$shard_id,
1109 : )
1110 : }};
1111 : }
1112 :
1113 : macro_rules! respond_error {
1114 : ($span:expr, $error:expr) => {{
1115 : let error = BatchedFeMessage::RespondError {
1116 : span: $span,
1117 : error: BatchedPageStreamError {
1118 : req: req.hdr,
1119 : err: $error,
1120 : },
1121 : };
1122 : Ok(Some(error))
1123 : }};
1124 : }
1125 :
1126 0 : let key = rel_block_to_key(req.rel, req.blkno);
1127 :
1128 0 : let res = timeline_handles
1129 0 : .get(tenant_id, timeline_id, ShardSelector::Page(key))
1130 0 : .await;
1131 :
1132 0 : let shard = match res {
1133 0 : Ok(tl) => tl,
1134 0 : Err(e) => {
1135 0 : let span = mkspan!(before shard routing);
1136 0 : match e {
1137 : GetActiveTimelineError::Tenant(GetActiveTenantError::NotFound(_)) => {
1138 : // We already know this tenant exists in general, because we resolved it at
1139 : // start of connection. Getting a NotFound here indicates that the shard containing
1140 : // the requested page is not present on this node: the client's knowledge of shard->pageserver
1141 : // mapping is out of date.
1142 : //
1143 : // Closing the connection by returning ``::Reconnect` has the side effect of rate-limiting above message, via
1144 : // client's reconnect backoff, as well as hopefully prompting the client to load its updated configuration
1145 : // and talk to a different pageserver.
1146 0 : MISROUTED_PAGESTREAM_REQUESTS.inc();
1147 0 : return respond_error!(
1148 0 : span,
1149 0 : PageStreamError::Reconnect(
1150 0 : "getpage@lsn request routed to wrong shard".into()
1151 0 : )
1152 : );
1153 : }
1154 0 : e => {
1155 0 : return respond_error!(span, e.into());
1156 : }
1157 : }
1158 : }
1159 : };
1160 :
1161 0 : let ctx = if shard.is_get_page_request_sampled() {
1162 0 : RequestContextBuilder::from(ctx)
1163 0 : .root_perf_span(|| {
1164 0 : info_span!(
1165 : target: PERF_TRACE_TARGET,
1166 : "GET_PAGE",
1167 : peer_addr = conn_perf_span_fields.peer_addr,
1168 : application_name = conn_perf_span_fields.application_name,
1169 : compute_mode = conn_perf_span_fields.compute_mode,
1170 : tenant_id = %tenant_id,
1171 0 : shard_id = %shard.get_shard_identity().shard_slug(),
1172 : timeline_id = %timeline_id,
1173 : lsn = %req.hdr.request_lsn,
1174 : not_modified_since_lsn = %req.hdr.not_modified_since,
1175 : request_id = %req.hdr.reqid,
1176 : key = %key,
1177 : )
1178 0 : })
1179 0 : .attached_child()
1180 : } else {
1181 0 : ctx.attached_child()
1182 : };
1183 :
1184 : // This ctx travels as part of the BatchedFeMessage through
1185 : // batching into the request handler.
1186 : // The request handler needs to do some per-request work
1187 : // (relsize check) before dispatching the batch as a single
1188 : // get_vectored call to the Timeline.
1189 : // This ctx will be used for the reslize check, whereas the
1190 : // get_vectored call will be a different ctx with separate
1191 : // perf span.
1192 0 : let ctx = ctx.with_scope_page_service_pagestream(&shard);
1193 :
1194 : // Similar game for this `span`: we funnel it through so that
1195 : // request handler log messages contain the request-specific fields.
1196 0 : let span = mkspan!(shard.tenant_shard_id.shard_slug());
1197 :
1198 0 : let timer = Self::record_op_start_and_throttle(
1199 0 : &shard,
1200 0 : metrics::SmgrQueryType::GetPageAtLsn,
1201 0 : received_at,
1202 : )
1203 0 : .maybe_perf_instrument(&ctx, |current_perf_span| {
1204 0 : info_span!(
1205 : target: PERF_TRACE_TARGET,
1206 0 : parent: current_perf_span,
1207 : "THROTTLE",
1208 : )
1209 0 : })
1210 0 : .await?;
1211 :
1212 0 : let applied_gc_cutoff_guard = shard.get_applied_gc_cutoff_lsn(); // hold guard
1213 : // We're holding the Handle
1214 0 : let effective_lsn = match Self::effective_request_lsn(
1215 0 : &shard,
1216 0 : shard.get_last_record_lsn(),
1217 0 : req.hdr.request_lsn,
1218 0 : req.hdr.not_modified_since,
1219 0 : &applied_gc_cutoff_guard,
1220 0 : ) {
1221 0 : Ok(lsn) => lsn,
1222 0 : Err(e) => {
1223 0 : return respond_error!(span, e);
1224 : }
1225 : };
1226 0 : let applied_gc_cutoff_guard = match hold_gc_cutoff_guard {
1227 0 : HoldAppliedGcCutoffGuard::Yes => Some(applied_gc_cutoff_guard),
1228 : HoldAppliedGcCutoffGuard::No => {
1229 0 : drop(applied_gc_cutoff_guard);
1230 0 : None
1231 : }
1232 : };
1233 :
1234 0 : let batch_wait_ctx = if ctx.has_perf_span() {
1235 : Some(
1236 0 : RequestContextBuilder::from(&ctx)
1237 0 : .perf_span(|crnt_perf_span| {
1238 0 : info_span!(
1239 : target: PERF_TRACE_TARGET,
1240 0 : parent: crnt_perf_span,
1241 : "WAIT_EXECUTOR",
1242 : )
1243 0 : })
1244 0 : .attached_child(),
1245 : )
1246 : } else {
1247 0 : None
1248 : };
1249 :
1250 : BatchedFeMessage::GetPage {
1251 0 : span,
1252 0 : shard: shard.downgrade(),
1253 0 : applied_gc_cutoff_guard,
1254 0 : pages: smallvec![BatchedGetPageRequest {
1255 0 : req,
1256 0 : timer,
1257 0 : lsn_range: LsnRange {
1258 0 : effective_lsn,
1259 0 : request_lsn: req.hdr.request_lsn
1260 0 : },
1261 0 : ctx,
1262 0 : batch_wait_ctx,
1263 0 : }],
1264 : // The executor grabs the batch when it becomes idle.
1265 : // Hence, [`GetPageBatchBreakReason::ExecutorSteal`] is the
1266 : // default reason for breaking the batch.
1267 0 : batch_break_reason: GetPageBatchBreakReason::ExecutorSteal,
1268 : }
1269 : }
1270 : #[cfg(feature = "testing")]
1271 0 : PagestreamFeMessage::Test(req) => {
1272 0 : let shard = timeline_handles
1273 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
1274 0 : .await?;
1275 0 : let span = tracing::info_span!(parent: &parent_span, "handle_test_request", shard_id = %shard.tenant_shard_id.shard_slug());
1276 0 : let timer = Self::record_op_start_and_throttle(
1277 0 : &shard,
1278 0 : metrics::SmgrQueryType::Test,
1279 0 : received_at,
1280 0 : )
1281 0 : .await?;
1282 0 : BatchedFeMessage::Test {
1283 0 : span,
1284 0 : shard: shard.downgrade(),
1285 0 : requests: vec![BatchedTestRequest { req, timer }],
1286 0 : }
1287 : }
1288 : };
1289 0 : Ok(Some(batched_msg))
1290 0 : }
1291 :
1292 : /// Starts a SmgrOpTimer at received_at and throttles the request.
1293 0 : async fn record_op_start_and_throttle(
1294 0 : shard: &Handle<TenantManagerTypes>,
1295 0 : op: metrics::SmgrQueryType,
1296 0 : received_at: Instant,
1297 0 : ) -> Result<SmgrOpTimer, QueryError> {
1298 : // It's important to start the smgr op metric recorder as early as possible
1299 : // so that the _started counters are incremented before we do
1300 : // any serious waiting, e.g., for throttle, batching, or actual request handling.
1301 0 : let mut timer = shard.query_metrics.start_smgr_op(op, received_at);
1302 0 : let now = Instant::now();
1303 0 : timer.observe_throttle_start(now);
1304 0 : let throttled = tokio::select! {
1305 0 : res = shard.pagestream_throttle.throttle(1, now) => res,
1306 0 : _ = shard.cancel.cancelled() => return Err(QueryError::Shutdown),
1307 : };
1308 0 : timer.observe_throttle_done(throttled);
1309 0 : Ok(timer)
1310 0 : }
1311 :
1312 : /// Post-condition: `batch` is Some()
1313 : #[instrument(skip_all, level = tracing::Level::TRACE)]
1314 : #[allow(clippy::boxed_local)]
1315 : fn pagestream_do_batch(
1316 : batching_strategy: PageServiceProtocolPipelinedBatchingStrategy,
1317 : max_batch_size: NonZeroUsize,
1318 : batch: &mut Result<BatchedFeMessage, QueryError>,
1319 : this_msg: Result<BatchedFeMessage, QueryError>,
1320 : ) -> Result<(), Result<BatchedFeMessage, QueryError>> {
1321 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
1322 :
1323 : let this_msg = match this_msg {
1324 : Ok(this_msg) => this_msg,
1325 : Err(e) => return Err(Err(e)),
1326 : };
1327 :
1328 : let eligible_batch = match batch {
1329 : Ok(b) => b,
1330 : Err(_) => {
1331 : return Err(Ok(this_msg));
1332 : }
1333 : };
1334 :
1335 : let batch_break =
1336 : eligible_batch.should_break_batch(&this_msg, max_batch_size, batching_strategy);
1337 :
1338 : match batch_break {
1339 : Some(reason) => {
1340 : if let BatchedFeMessage::GetPage {
1341 : batch_break_reason, ..
1342 : } = eligible_batch
1343 : {
1344 : *batch_break_reason = reason;
1345 : }
1346 :
1347 : Err(Ok(this_msg))
1348 : }
1349 : None => {
1350 : // ok to batch
1351 : match (eligible_batch, this_msg) {
1352 : (
1353 : BatchedFeMessage::GetPage {
1354 : pages: accum_pages,
1355 : applied_gc_cutoff_guard: accum_applied_gc_cutoff_guard,
1356 : ..
1357 : },
1358 : BatchedFeMessage::GetPage {
1359 : pages: this_pages,
1360 : applied_gc_cutoff_guard: this_applied_gc_cutoff_guard,
1361 : ..
1362 : },
1363 : ) => {
1364 : accum_pages.extend(this_pages);
1365 : // the minimum of the two guards will keep data for both alive
1366 : match (&accum_applied_gc_cutoff_guard, this_applied_gc_cutoff_guard) {
1367 : (None, None) => (),
1368 : (None, Some(this)) => *accum_applied_gc_cutoff_guard = Some(this),
1369 : (Some(_), None) => (),
1370 : (Some(accum), Some(this)) => {
1371 : if **accum > *this {
1372 : *accum_applied_gc_cutoff_guard = Some(this);
1373 : }
1374 : }
1375 : };
1376 : Ok(())
1377 : }
1378 : #[cfg(feature = "testing")]
1379 : (
1380 : BatchedFeMessage::Test {
1381 : requests: accum_requests,
1382 : ..
1383 : },
1384 : BatchedFeMessage::Test {
1385 : requests: this_requests,
1386 : ..
1387 : },
1388 : ) => {
1389 : accum_requests.extend(this_requests);
1390 : Ok(())
1391 : }
1392 : // Shape guaranteed by [`BatchedFeMessage::should_break_batch`]
1393 : _ => unreachable!(),
1394 : }
1395 : }
1396 : }
1397 : }
1398 :
1399 0 : #[instrument(level = tracing::Level::DEBUG, skip_all)]
1400 : async fn pagestream_handle_batched_message<IO>(
1401 : &mut self,
1402 : pgb_writer: &mut PostgresBackend<IO>,
1403 : batch: BatchedFeMessage,
1404 : io_concurrency: IoConcurrency,
1405 : cancel: &CancellationToken,
1406 : protocol_version: PagestreamProtocolVersion,
1407 : ctx: &RequestContext,
1408 : ) -> Result<(), QueryError>
1409 : where
1410 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
1411 : {
1412 : let started_at = Instant::now();
1413 : let batch = {
1414 : let mut batch = batch;
1415 : batch.observe_execution_start(started_at);
1416 : batch
1417 : };
1418 :
1419 : // Dispatch the batch to the appropriate request handler.
1420 : let log_slow_name = batch.as_static_str();
1421 : let (mut handler_results, span) = {
1422 : // TODO: we unfortunately have to pin the future on the heap, since GetPage futures are huge and
1423 : // won't fit on the stack.
1424 : let mut boxpinned = Box::pin(Self::pagestream_dispatch_batched_message(
1425 : batch,
1426 : io_concurrency,
1427 : ctx,
1428 : ));
1429 : log_slow(
1430 : log_slow_name,
1431 : LOG_SLOW_GETPAGE_THRESHOLD,
1432 : boxpinned.as_mut(),
1433 : )
1434 : .await?
1435 : };
1436 :
1437 : // We purposefully don't count flush time into the smgr operation timer.
1438 : //
1439 : // The reason is that current compute client will not perform protocol processing
1440 : // if the postgres backend process is doing things other than `->smgr_read()`.
1441 : // This is especially the case for prefetch.
1442 : //
1443 : // If the compute doesn't read from the connection, eventually TCP will backpressure
1444 : // all the way into our flush call below.
1445 : //
1446 : // The timer's underlying metric is used for a storage-internal latency SLO and
1447 : // we don't want to include latency in it that we can't control.
1448 : // And as pointed out above, in this case, we don't control the time that flush will take.
1449 : //
1450 : // We put each response in the batch onto the wire in a separate pgb_writer.flush()
1451 : // call, which (all unmeasured) adds syscall overhead but reduces time to first byte
1452 : // and avoids building up a "giant" contiguous userspace buffer to hold the entire response.
1453 : // TODO: vectored socket IO would be great, but pgb_writer doesn't support that.
1454 : let flush_timers = {
1455 : let flushing_start_time = Instant::now();
1456 : let mut flush_timers = Vec::with_capacity(handler_results.len());
1457 : for handler_result in &mut handler_results {
1458 : let flush_timer = match handler_result {
1459 : Ok((_response, timer, _ctx)) => Some(
1460 : timer
1461 : .observe_execution_end(flushing_start_time)
1462 : .expect("we are the first caller"),
1463 : ),
1464 : Err(_) => {
1465 : // TODO: measure errors
1466 : None
1467 : }
1468 : };
1469 : flush_timers.push(flush_timer);
1470 : }
1471 : assert_eq!(flush_timers.len(), handler_results.len());
1472 : flush_timers
1473 : };
1474 :
1475 : // Map handler result to protocol behavior.
1476 : // Some handler errors cause exit from pagestream protocol.
1477 : // Other handler errors are sent back as an error message and we stay in pagestream protocol.
1478 : for (handler_result, flushing_timer) in handler_results.into_iter().zip(flush_timers) {
1479 : let (response_msg, ctx) = match handler_result {
1480 : Err(e) => match &e.err {
1481 : PageStreamError::Shutdown => {
1482 : // BEGIN HADRON
1483 : PAGESTREAM_HANDLER_RESULTS_TOTAL
1484 : .with_label_values(&[metrics::PAGESTREAM_HANDLER_OUTCOME_OTHER_ERROR])
1485 : .inc();
1486 : // END HADRON
1487 :
1488 : // If we fail to fulfil a request during shutdown, which may be _because_ of
1489 : // shutdown, then do not send the error to the client. Instead just drop the
1490 : // connection.
1491 0 : span.in_scope(|| info!("dropping connection due to shutdown"));
1492 : return Err(QueryError::Shutdown);
1493 : }
1494 : PageStreamError::Reconnect(_reason) => {
1495 0 : span.in_scope(|| {
1496 : // BEGIN HADRON
1497 : // We can get here because the compute node is pointing at the wrong PS. We
1498 : // already have a metric to keep track of this so suppressing this log to
1499 : // reduce log spam. The information in this log message is not going to be that
1500 : // helpful given the volume of logs that can be generated.
1501 : // info!("handler requested reconnect: {reason}")
1502 : // END HADRON
1503 0 : });
1504 : // BEGIN HADRON
1505 : PAGESTREAM_HANDLER_RESULTS_TOTAL
1506 : .with_label_values(&[
1507 : metrics::PAGESTREAM_HANDLER_OUTCOME_INTERNAL_ERROR,
1508 : ])
1509 : .inc();
1510 : // END HADRON
1511 : return Err(QueryError::Reconnect);
1512 : }
1513 : PageStreamError::Read(_)
1514 : | PageStreamError::LsnTimeout(_)
1515 : | PageStreamError::NotFound(_)
1516 : | PageStreamError::BadRequest(_) => {
1517 : // BEGIN HADRON
1518 : if let PageStreamError::Read(_) | PageStreamError::LsnTimeout(_) = &e.err {
1519 : PAGESTREAM_HANDLER_RESULTS_TOTAL
1520 : .with_label_values(&[
1521 : metrics::PAGESTREAM_HANDLER_OUTCOME_INTERNAL_ERROR,
1522 : ])
1523 : .inc();
1524 : } else {
1525 : PAGESTREAM_HANDLER_RESULTS_TOTAL
1526 : .with_label_values(&[
1527 : metrics::PAGESTREAM_HANDLER_OUTCOME_OTHER_ERROR,
1528 : ])
1529 : .inc();
1530 : }
1531 : // END HADRON
1532 :
1533 : // print the all details to the log with {:#}, but for the client the
1534 : // error message is enough. Do not log if shutting down, as the anyhow::Error
1535 : // here includes cancellation which is not an error.
1536 : let full = utils::error::report_compact_sources(&e.err);
1537 0 : span.in_scope(|| {
1538 0 : error!("error reading relation or page version: {full:#}")
1539 0 : });
1540 :
1541 : (
1542 : PagestreamBeMessage::Error(PagestreamErrorResponse {
1543 : req: e.req,
1544 : message: e.err.to_string(),
1545 : }),
1546 : None,
1547 : )
1548 : }
1549 : },
1550 : Ok((response_msg, _op_timer_already_observed, ctx)) => {
1551 : // BEGIN HADRON
1552 : PAGESTREAM_HANDLER_RESULTS_TOTAL
1553 : .with_label_values(&[metrics::PAGESTREAM_HANDLER_OUTCOME_SUCCESS])
1554 : .inc();
1555 : // END HADRON
1556 :
1557 : (response_msg, Some(ctx))
1558 : }
1559 : };
1560 :
1561 0 : let ctx = ctx.map(|req_ctx| {
1562 0 : RequestContextBuilder::from(&req_ctx)
1563 0 : .perf_span(|crnt_perf_span| {
1564 0 : info_span!(
1565 : target: PERF_TRACE_TARGET,
1566 0 : parent: crnt_perf_span,
1567 : "FLUSH_RESPONSE",
1568 : )
1569 0 : })
1570 0 : .attached_child()
1571 0 : });
1572 :
1573 : //
1574 : // marshal & transmit response message
1575 : //
1576 :
1577 : pgb_writer.write_message_noflush(&BeMessage::CopyData(
1578 : &response_msg.serialize(protocol_version),
1579 : ))?;
1580 :
1581 : failpoint_support::sleep_millis_async!("before-pagestream-msg-flush", cancel);
1582 :
1583 : // what we want to do
1584 : let socket_fd = pgb_writer.socket_fd;
1585 : let flush_fut = pgb_writer.flush();
1586 : // metric for how long flushing takes
1587 : let flush_fut = match flushing_timer {
1588 : Some(flushing_timer) => futures::future::Either::Left(flushing_timer.measure(
1589 : Instant::now(),
1590 : flush_fut,
1591 : socket_fd,
1592 : )),
1593 : None => futures::future::Either::Right(flush_fut),
1594 : };
1595 :
1596 : let flush_fut = if let Some(req_ctx) = ctx.as_ref() {
1597 : futures::future::Either::Left(
1598 0 : flush_fut.maybe_perf_instrument(req_ctx, |current_perf_span| {
1599 0 : current_perf_span.clone()
1600 0 : }),
1601 : )
1602 : } else {
1603 : futures::future::Either::Right(flush_fut)
1604 : };
1605 :
1606 : // do it while respecting cancellation
1607 0 : let _: () = async move {
1608 0 : tokio::select! {
1609 : biased;
1610 0 : _ = cancel.cancelled() => {
1611 : // We were requested to shut down.
1612 0 : info!("shutdown request received in page handler");
1613 0 : return Err(QueryError::Shutdown)
1614 : }
1615 0 : res = flush_fut => {
1616 0 : res?;
1617 : }
1618 : }
1619 0 : Ok(())
1620 0 : }
1621 : .await?;
1622 : }
1623 : Ok(())
1624 : }
1625 :
1626 : /// Helper which dispatches a batched message to the appropriate handler.
1627 : /// Returns a vec of results, along with the extracted trace span.
1628 0 : async fn pagestream_dispatch_batched_message(
1629 0 : batch: BatchedFeMessage,
1630 0 : io_concurrency: IoConcurrency,
1631 0 : ctx: &RequestContext,
1632 0 : ) -> Result<
1633 0 : (
1634 0 : Vec<Result<(PagestreamBeMessage, SmgrOpTimer, RequestContext), BatchedPageStreamError>>,
1635 0 : Span,
1636 0 : ),
1637 0 : QueryError,
1638 0 : > {
1639 : macro_rules! upgrade_handle_and_set_context {
1640 : ($shard:ident) => {{
1641 : let weak_handle = &$shard;
1642 : let handle = weak_handle.upgrade()?;
1643 : let ctx = ctx.with_scope_page_service_pagestream(&handle);
1644 : (handle, ctx)
1645 : }};
1646 : }
1647 0 : Ok(match batch {
1648 : BatchedFeMessage::Exists {
1649 0 : span,
1650 0 : timer,
1651 0 : shard,
1652 0 : req,
1653 : } => {
1654 0 : fail::fail_point!("ps::handle-pagerequest-message::exists");
1655 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1656 : (
1657 0 : vec![
1658 0 : Self::handle_get_rel_exists_request(&shard, &req, &ctx)
1659 0 : .instrument(span.clone())
1660 0 : .await
1661 0 : .map(|msg| (PagestreamBeMessage::Exists(msg), timer, ctx))
1662 0 : .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
1663 : ],
1664 0 : span,
1665 : )
1666 : }
1667 : BatchedFeMessage::Nblocks {
1668 0 : span,
1669 0 : timer,
1670 0 : shard,
1671 0 : req,
1672 : } => {
1673 0 : fail::fail_point!("ps::handle-pagerequest-message::nblocks");
1674 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1675 : (
1676 0 : vec![
1677 0 : Self::handle_get_nblocks_request(&shard, &req, false, &ctx)
1678 0 : .instrument(span.clone())
1679 0 : .await
1680 0 : .map(|msg| msg.expect("allow_missing=false"))
1681 0 : .map(|msg| (PagestreamBeMessage::Nblocks(msg), timer, ctx))
1682 0 : .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
1683 : ],
1684 0 : span,
1685 : )
1686 : }
1687 : BatchedFeMessage::GetPage {
1688 0 : span,
1689 0 : shard,
1690 0 : applied_gc_cutoff_guard,
1691 0 : pages,
1692 0 : batch_break_reason,
1693 : } => {
1694 0 : fail::fail_point!("ps::handle-pagerequest-message::getpage");
1695 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1696 : (
1697 : {
1698 0 : let npages = pages.len();
1699 0 : trace!(npages, "handling getpage request");
1700 0 : let res = Self::handle_get_page_at_lsn_request_batched(
1701 0 : &shard,
1702 0 : pages,
1703 0 : io_concurrency,
1704 0 : batch_break_reason,
1705 0 : &ctx,
1706 0 : )
1707 0 : .instrument(span.clone())
1708 0 : .await;
1709 0 : assert_eq!(res.len(), npages);
1710 0 : drop(applied_gc_cutoff_guard);
1711 0 : res
1712 : },
1713 0 : span,
1714 : )
1715 : }
1716 : BatchedFeMessage::DbSize {
1717 0 : span,
1718 0 : timer,
1719 0 : shard,
1720 0 : req,
1721 : } => {
1722 0 : fail::fail_point!("ps::handle-pagerequest-message::dbsize");
1723 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1724 : (
1725 0 : vec![
1726 0 : Self::handle_db_size_request(&shard, &req, &ctx)
1727 0 : .instrument(span.clone())
1728 0 : .await
1729 0 : .map(|msg| (PagestreamBeMessage::DbSize(msg), timer, ctx))
1730 0 : .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
1731 : ],
1732 0 : span,
1733 : )
1734 : }
1735 : BatchedFeMessage::GetSlruSegment {
1736 0 : span,
1737 0 : timer,
1738 0 : shard,
1739 0 : req,
1740 : } => {
1741 0 : fail::fail_point!("ps::handle-pagerequest-message::slrusegment");
1742 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1743 : (
1744 0 : vec![
1745 0 : Self::handle_get_slru_segment_request(&shard, &req, &ctx)
1746 0 : .instrument(span.clone())
1747 0 : .await
1748 0 : .map(|msg| (PagestreamBeMessage::GetSlruSegment(msg), timer, ctx))
1749 0 : .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
1750 : ],
1751 0 : span,
1752 : )
1753 : }
1754 : #[cfg(feature = "testing")]
1755 : BatchedFeMessage::Test {
1756 0 : span,
1757 0 : shard,
1758 0 : requests,
1759 : } => {
1760 0 : fail::fail_point!("ps::handle-pagerequest-message::test");
1761 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1762 : (
1763 : {
1764 0 : let npages = requests.len();
1765 0 : trace!(npages, "handling getpage request");
1766 0 : let res = Self::handle_test_request_batch(&shard, requests, &ctx)
1767 0 : .instrument(span.clone())
1768 0 : .await;
1769 0 : assert_eq!(res.len(), npages);
1770 0 : res
1771 : },
1772 0 : span,
1773 : )
1774 : }
1775 0 : BatchedFeMessage::RespondError { span, error } => {
1776 : // We've already decided to respond with an error, so we don't need to
1777 : // call the handler.
1778 0 : (vec![Err(error)], span)
1779 : }
1780 : })
1781 0 : }
1782 :
1783 : /// Pagestream sub-protocol handler.
1784 : ///
1785 : /// It is a simple request-response protocol inside a COPYBOTH session.
1786 : ///
1787 : /// # Coding Discipline
1788 : ///
1789 : /// Coding discipline within this function: all interaction with the `pgb` connection
1790 : /// needs to be sensitive to connection shutdown, currently signalled via [`Self::cancel`].
1791 : /// This is so that we can shutdown page_service quickly.
1792 : #[instrument(skip_all, fields(hold_gc_cutoff_guard))]
1793 : async fn handle_pagerequests<IO>(
1794 : &mut self,
1795 : pgb: &mut PostgresBackend<IO>,
1796 : tenant_id: TenantId,
1797 : timeline_id: TimelineId,
1798 : protocol_version: PagestreamProtocolVersion,
1799 : ctx: RequestContext,
1800 : ) -> Result<(), QueryError>
1801 : where
1802 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
1803 : {
1804 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
1805 :
1806 : // switch client to COPYBOTH
1807 : pgb.write_message_noflush(&BeMessage::CopyBothResponse)?;
1808 : tokio::select! {
1809 : biased;
1810 : _ = self.cancel.cancelled() => {
1811 : return Err(QueryError::Shutdown)
1812 : }
1813 : res = pgb.flush() => {
1814 : res?;
1815 : }
1816 : }
1817 :
1818 : let io_concurrency = IoConcurrency::spawn_from_conf(
1819 : self.get_vectored_concurrent_io,
1820 : match self.gate_guard.try_clone() {
1821 : Ok(guard) => guard,
1822 : Err(_) => {
1823 : info!("shutdown request received in page handler");
1824 : return Err(QueryError::Shutdown);
1825 : }
1826 : },
1827 : );
1828 :
1829 : let pgb_reader = pgb
1830 : .split()
1831 : .context("implementation error: split pgb into reader and writer")?;
1832 :
1833 : let timeline_handles = self
1834 : .timeline_handles
1835 : .take()
1836 : .expect("implementation error: timeline_handles should not be locked");
1837 :
1838 : // Evaluate the expensive feature resolver check once per pagestream subprotocol handling
1839 : // instead of once per GetPage request. This is shared between pipelined and serial paths.
1840 : let hold_gc_cutoff_guard = if cfg!(test) || cfg!(feature = "testing") {
1841 : HoldAppliedGcCutoffGuard::Yes
1842 : } else {
1843 : // Use the global feature resolver with the tenant ID directly, avoiding the need
1844 : // to get a timeline/shard which might not be available on this pageserver node.
1845 : let empty_properties = std::collections::HashMap::new();
1846 : match self.feature_resolver.evaluate_boolean(
1847 : "page-service-getpage-hold-applied-gc-cutoff-guard",
1848 : tenant_id,
1849 : &empty_properties,
1850 : ) {
1851 : Ok(()) => HoldAppliedGcCutoffGuard::Yes,
1852 : Err(_) => HoldAppliedGcCutoffGuard::No,
1853 : }
1854 : };
1855 : // record it in the span of handle_pagerequests so that both the request_span
1856 : // and the pipeline implementation spans contains the field.
1857 : Span::current().record(
1858 : "hold_gc_cutoff_guard",
1859 : tracing::field::debug(&hold_gc_cutoff_guard),
1860 : );
1861 :
1862 : let request_span = info_span!("request");
1863 : let ((pgb_reader, timeline_handles), result) = match self.pipelining_config.clone() {
1864 : PageServicePipeliningConfig::Pipelined(pipelining_config) => {
1865 : self.handle_pagerequests_pipelined(
1866 : pgb,
1867 : pgb_reader,
1868 : tenant_id,
1869 : timeline_id,
1870 : timeline_handles,
1871 : request_span,
1872 : pipelining_config,
1873 : protocol_version,
1874 : io_concurrency,
1875 : hold_gc_cutoff_guard,
1876 : &ctx,
1877 : )
1878 : .await
1879 : }
1880 : PageServicePipeliningConfig::Serial => {
1881 : self.handle_pagerequests_serial(
1882 : pgb,
1883 : pgb_reader,
1884 : tenant_id,
1885 : timeline_id,
1886 : timeline_handles,
1887 : request_span,
1888 : protocol_version,
1889 : io_concurrency,
1890 : hold_gc_cutoff_guard,
1891 : &ctx,
1892 : )
1893 : .await
1894 : }
1895 : };
1896 :
1897 : debug!("pagestream subprotocol shut down cleanly");
1898 :
1899 : pgb.unsplit(pgb_reader)
1900 : .context("implementation error: unsplit pgb")?;
1901 :
1902 : let replaced = self.timeline_handles.replace(timeline_handles);
1903 : assert!(replaced.is_none());
1904 :
1905 : result
1906 : }
1907 :
1908 : #[allow(clippy::too_many_arguments)]
1909 0 : async fn handle_pagerequests_serial<IO>(
1910 0 : &mut self,
1911 0 : pgb_writer: &mut PostgresBackend<IO>,
1912 0 : mut pgb_reader: PostgresBackendReader<IO>,
1913 0 : tenant_id: TenantId,
1914 0 : timeline_id: TimelineId,
1915 0 : mut timeline_handles: TimelineHandles,
1916 0 : request_span: Span,
1917 0 : protocol_version: PagestreamProtocolVersion,
1918 0 : io_concurrency: IoConcurrency,
1919 0 : hold_gc_cutoff_guard: HoldAppliedGcCutoffGuard,
1920 0 : ctx: &RequestContext,
1921 0 : ) -> (
1922 0 : (PostgresBackendReader<IO>, TimelineHandles),
1923 0 : Result<(), QueryError>,
1924 0 : )
1925 0 : where
1926 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
1927 0 : {
1928 0 : let cancel = self.cancel.clone();
1929 :
1930 0 : let err = loop {
1931 0 : let msg = Self::pagestream_read_message(
1932 0 : &mut pgb_reader,
1933 0 : tenant_id,
1934 0 : timeline_id,
1935 0 : &mut timeline_handles,
1936 0 : &self.perf_span_fields,
1937 0 : &cancel,
1938 0 : ctx,
1939 0 : protocol_version,
1940 0 : request_span.clone(),
1941 0 : hold_gc_cutoff_guard,
1942 0 : )
1943 0 : .await;
1944 0 : let msg = match msg {
1945 0 : Ok(msg) => msg,
1946 0 : Err(e) => break e,
1947 : };
1948 0 : let msg = match msg {
1949 0 : Some(msg) => msg,
1950 : None => {
1951 0 : debug!("pagestream subprotocol end observed");
1952 0 : return ((pgb_reader, timeline_handles), Ok(()));
1953 : }
1954 : };
1955 :
1956 0 : let result = self
1957 0 : .pagestream_handle_batched_message(
1958 0 : pgb_writer,
1959 0 : msg,
1960 0 : io_concurrency.clone(),
1961 0 : &cancel,
1962 0 : protocol_version,
1963 0 : ctx,
1964 0 : )
1965 0 : .await;
1966 0 : match result {
1967 0 : Ok(()) => {}
1968 0 : Err(e) => break e,
1969 : }
1970 : };
1971 0 : ((pgb_reader, timeline_handles), Err(err))
1972 0 : }
1973 :
1974 : /// # Cancel-Safety
1975 : ///
1976 : /// May leak tokio tasks if not polled to completion.
1977 : #[allow(clippy::too_many_arguments)]
1978 0 : async fn handle_pagerequests_pipelined<IO>(
1979 0 : &mut self,
1980 0 : pgb_writer: &mut PostgresBackend<IO>,
1981 0 : pgb_reader: PostgresBackendReader<IO>,
1982 0 : tenant_id: TenantId,
1983 0 : timeline_id: TimelineId,
1984 0 : mut timeline_handles: TimelineHandles,
1985 0 : request_span: Span,
1986 0 : pipelining_config: PageServicePipeliningConfigPipelined,
1987 0 : protocol_version: PagestreamProtocolVersion,
1988 0 : io_concurrency: IoConcurrency,
1989 0 : hold_gc_cutoff_guard: HoldAppliedGcCutoffGuard,
1990 0 : ctx: &RequestContext,
1991 0 : ) -> (
1992 0 : (PostgresBackendReader<IO>, TimelineHandles),
1993 0 : Result<(), QueryError>,
1994 0 : )
1995 0 : where
1996 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
1997 0 : {
1998 : //
1999 : // Pipelined pagestream handling consists of
2000 : // - a Batcher that reads requests off the wire and
2001 : // and batches them if possible,
2002 : // - an Executor that processes the batched requests.
2003 : //
2004 : // The batch is built up inside an `spsc_fold` channel,
2005 : // shared betwen Batcher (Sender) and Executor (Receiver).
2006 : //
2007 : // The Batcher continously folds client requests into the batch,
2008 : // while the Executor can at any time take out what's in the batch
2009 : // in order to process it.
2010 : // This means the next batch builds up while the Executor
2011 : // executes the last batch.
2012 : //
2013 : // CANCELLATION
2014 : //
2015 : // We run both Batcher and Executor futures to completion before
2016 : // returning from this function.
2017 : //
2018 : // If Executor exits first, it signals cancellation to the Batcher
2019 : // via a CancellationToken that is child of `self.cancel`.
2020 : // If Batcher exits first, it signals cancellation to the Executor
2021 : // by dropping the spsc_fold channel Sender.
2022 : //
2023 : // CLEAN SHUTDOWN
2024 : //
2025 : // Clean shutdown means that the client ends the COPYBOTH session.
2026 : // In response to such a client message, the Batcher exits.
2027 : // The Executor continues to run, draining the spsc_fold channel.
2028 : // Once drained, the spsc_fold recv will fail with a distinct error
2029 : // indicating that the sender disconnected.
2030 : // The Executor exits with Ok(()) in response to that error.
2031 : //
2032 : // Server initiated shutdown is not clean shutdown, but instead
2033 : // is an error Err(QueryError::Shutdown) that is propagated through
2034 : // error propagation.
2035 : //
2036 : // ERROR PROPAGATION
2037 : //
2038 : // When the Batcher encounter an error, it sends it as a value
2039 : // through the spsc_fold channel and exits afterwards.
2040 : // When the Executor observes such an error in the channel,
2041 : // it exits returning that error value.
2042 : //
2043 : // This design ensures that the Executor stage will still process
2044 : // the batch that was in flight when the Batcher encountered an error,
2045 : // thereby beahving identical to a serial implementation.
2046 :
2047 : let PageServicePipeliningConfigPipelined {
2048 0 : max_batch_size,
2049 0 : execution,
2050 0 : batching: batching_strategy,
2051 0 : } = pipelining_config;
2052 :
2053 : // Macro to _define_ a pipeline stage.
2054 : macro_rules! pipeline_stage {
2055 : ($name:literal, $cancel:expr, $make_fut:expr) => {{
2056 : let cancel: CancellationToken = $cancel;
2057 : let stage_fut = $make_fut(cancel.clone());
2058 0 : async move {
2059 0 : scopeguard::defer! {
2060 : debug!("exiting");
2061 : }
2062 0 : timed_after_cancellation(stage_fut, $name, Duration::from_millis(100), &cancel)
2063 0 : .await
2064 0 : }
2065 : .instrument(tracing::info_span!($name))
2066 : }};
2067 : }
2068 :
2069 : //
2070 : // Batcher
2071 : //
2072 :
2073 0 : let perf_span_fields = self.perf_span_fields.clone();
2074 :
2075 0 : let cancel_batcher = self.cancel.child_token();
2076 0 : let (mut batch_tx, mut batch_rx) = spsc_fold::channel();
2077 0 : let batcher = pipeline_stage!("batcher", cancel_batcher.clone(), move |cancel_batcher| {
2078 0 : let ctx = ctx.attached_child();
2079 0 : async move {
2080 0 : let mut pgb_reader = pgb_reader;
2081 0 : let mut exit = false;
2082 0 : while !exit {
2083 0 : let read_res = Self::pagestream_read_message(
2084 0 : &mut pgb_reader,
2085 0 : tenant_id,
2086 0 : timeline_id,
2087 0 : &mut timeline_handles,
2088 0 : &perf_span_fields,
2089 0 : &cancel_batcher,
2090 0 : &ctx,
2091 0 : protocol_version,
2092 0 : request_span.clone(),
2093 0 : hold_gc_cutoff_guard,
2094 0 : )
2095 0 : .await;
2096 0 : let Some(read_res) = read_res.transpose() else {
2097 0 : debug!("client-initiated shutdown");
2098 0 : break;
2099 : };
2100 0 : exit |= read_res.is_err();
2101 0 : let could_send = batch_tx
2102 0 : .send(read_res, |batch, res| {
2103 0 : Self::pagestream_do_batch(batching_strategy, max_batch_size, batch, res)
2104 0 : })
2105 0 : .await;
2106 0 : exit |= could_send.is_err();
2107 : }
2108 0 : (pgb_reader, timeline_handles)
2109 0 : }
2110 0 : });
2111 :
2112 : //
2113 : // Executor
2114 : //
2115 :
2116 0 : let executor = pipeline_stage!("executor", self.cancel.clone(), move |cancel| {
2117 0 : let ctx = ctx.attached_child();
2118 0 : async move {
2119 0 : let _cancel_batcher = cancel_batcher.drop_guard();
2120 : loop {
2121 0 : let maybe_batch = batch_rx.recv().await;
2122 0 : let batch = match maybe_batch {
2123 0 : Ok(batch) => batch,
2124 : Err(spsc_fold::RecvError::SenderGone) => {
2125 0 : debug!("upstream gone");
2126 0 : return Ok(());
2127 : }
2128 : };
2129 0 : let mut batch = match batch {
2130 0 : Ok(batch) => batch,
2131 0 : Err(e) => {
2132 0 : return Err(e);
2133 : }
2134 : };
2135 :
2136 : if let BatchedFeMessage::GetPage {
2137 0 : pages,
2138 : span: _,
2139 : shard: _,
2140 : applied_gc_cutoff_guard: _,
2141 : batch_break_reason: _,
2142 0 : } = &mut batch
2143 : {
2144 0 : for req in pages {
2145 0 : req.batch_wait_ctx.take();
2146 0 : }
2147 0 : }
2148 :
2149 0 : self.pagestream_handle_batched_message(
2150 0 : pgb_writer,
2151 0 : batch,
2152 0 : io_concurrency.clone(),
2153 0 : &cancel,
2154 0 : protocol_version,
2155 0 : &ctx,
2156 0 : )
2157 0 : .await?;
2158 : }
2159 0 : }
2160 0 : });
2161 :
2162 : //
2163 : // Execute the stages.
2164 : //
2165 :
2166 0 : match execution {
2167 : PageServiceProtocolPipelinedExecutionStrategy::ConcurrentFutures => {
2168 0 : tokio::join!(batcher, executor)
2169 : }
2170 : PageServiceProtocolPipelinedExecutionStrategy::Tasks => {
2171 : // These tasks are not tracked anywhere.
2172 0 : let read_messages_task = tokio::spawn(batcher);
2173 0 : let (read_messages_task_res, executor_res_) =
2174 0 : tokio::join!(read_messages_task, executor,);
2175 0 : (
2176 0 : read_messages_task_res.expect("propagated panic from read_messages"),
2177 0 : executor_res_,
2178 0 : )
2179 : }
2180 : }
2181 0 : }
2182 :
2183 : /// Helper function to handle the LSN from client request.
2184 : ///
2185 : /// Each GetPage (and Exists and Nblocks) request includes information about
2186 : /// which version of the page is being requested. The primary compute node
2187 : /// will always request the latest page version, by setting 'request_lsn' to
2188 : /// the last inserted or flushed WAL position, while a standby will request
2189 : /// a version at the LSN that it's currently caught up to.
2190 : ///
2191 : /// In either case, if the page server hasn't received the WAL up to the
2192 : /// requested LSN yet, we will wait for it to arrive. The return value is
2193 : /// the LSN that should be used to look up the page versions.
2194 : ///
2195 : /// In addition to the request LSN, each request carries another LSN,
2196 : /// 'not_modified_since', which is a hint to the pageserver that the client
2197 : /// knows that the page has not been modified between 'not_modified_since'
2198 : /// and the request LSN. This allows skipping the wait, as long as the WAL
2199 : /// up to 'not_modified_since' has arrived. If the client doesn't have any
2200 : /// information about when the page was modified, it will use
2201 : /// not_modified_since == lsn. If the client lies and sends a too low
2202 : /// not_modified_hint such that there are in fact later page versions, the
2203 : /// behavior is undefined: the pageserver may return any of the page versions
2204 : /// or an error.
2205 0 : async fn wait_or_get_last_lsn(
2206 0 : timeline: &Timeline,
2207 0 : request_lsn: Lsn,
2208 0 : not_modified_since: Lsn,
2209 0 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
2210 0 : ctx: &RequestContext,
2211 0 : ) -> Result<Lsn, PageStreamError> {
2212 0 : let last_record_lsn = timeline.get_last_record_lsn();
2213 0 : let effective_request_lsn = Self::effective_request_lsn(
2214 0 : timeline,
2215 0 : last_record_lsn,
2216 0 : request_lsn,
2217 0 : not_modified_since,
2218 0 : latest_gc_cutoff_lsn,
2219 0 : )?;
2220 :
2221 0 : if effective_request_lsn > last_record_lsn {
2222 0 : timeline
2223 0 : .wait_lsn(
2224 0 : not_modified_since,
2225 0 : crate::tenant::timeline::WaitLsnWaiter::PageService,
2226 0 : timeline::WaitLsnTimeout::Default,
2227 0 : ctx,
2228 0 : )
2229 0 : .await?;
2230 :
2231 : // Since we waited for 'effective_request_lsn' to arrive, that is now the last
2232 : // record LSN. (Or close enough for our purposes; the last-record LSN can
2233 : // advance immediately after we return anyway)
2234 0 : }
2235 :
2236 0 : Ok(effective_request_lsn)
2237 0 : }
2238 :
2239 0 : fn effective_request_lsn(
2240 0 : timeline: &Timeline,
2241 0 : last_record_lsn: Lsn,
2242 0 : request_lsn: Lsn,
2243 0 : not_modified_since: Lsn,
2244 0 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
2245 0 : ) -> Result<Lsn, PageStreamError> {
2246 : // Sanity check the request
2247 0 : if request_lsn < not_modified_since {
2248 0 : return Err(PageStreamError::BadRequest(
2249 0 : format!(
2250 0 : "invalid request with request LSN {request_lsn} and not_modified_since {not_modified_since}",
2251 0 : )
2252 0 : .into(),
2253 0 : ));
2254 0 : }
2255 :
2256 : // Check explicitly for INVALID just to get a less scary error message if the request is obviously bogus
2257 0 : if request_lsn == Lsn::INVALID {
2258 0 : return Err(PageStreamError::BadRequest(
2259 0 : "invalid LSN(0) in request".into(),
2260 0 : ));
2261 0 : }
2262 :
2263 : // Clients should only read from recent LSNs on their timeline, or from locations holding an LSN lease.
2264 : //
2265 : // We may have older data available, but we make a best effort to detect this case and return an error,
2266 : // to distinguish a misbehaving client (asking for old LSN) from a storage issue (data missing at a legitimate LSN).
2267 0 : if request_lsn < **latest_gc_cutoff_lsn && !timeline.is_gc_blocked_by_lsn_lease_deadline() {
2268 0 : let gc_info = &timeline.gc_info.read().unwrap();
2269 0 : if !gc_info.lsn_covered_by_lease(request_lsn) {
2270 0 : return Err(
2271 0 : PageStreamError::BadRequest(format!(
2272 0 : "tried to request a page version that was garbage collected. requested at {} gc cutoff {}",
2273 0 : request_lsn, **latest_gc_cutoff_lsn
2274 0 : ).into())
2275 0 : );
2276 0 : }
2277 0 : }
2278 :
2279 0 : if not_modified_since > last_record_lsn {
2280 0 : Ok(not_modified_since)
2281 : } else {
2282 : // It might be better to use max(not_modified_since, latest_gc_cutoff_lsn)
2283 : // here instead. That would give the same result, since we know that there
2284 : // haven't been any modifications since 'not_modified_since'. Using an older
2285 : // LSN might be faster, because that could allow skipping recent layers when
2286 : // finding the page. However, we have historically used 'last_record_lsn', so
2287 : // stick to that for now.
2288 0 : Ok(std::cmp::min(last_record_lsn, request_lsn))
2289 : }
2290 0 : }
2291 :
2292 : /// Handles the lsn lease request.
2293 : /// If a lease cannot be obtained, the client will receive NULL.
2294 : #[instrument(skip_all, fields(shard_id, %lsn))]
2295 : async fn handle_make_lsn_lease<IO>(
2296 : &mut self,
2297 : pgb: &mut PostgresBackend<IO>,
2298 : tenant_shard_id: TenantShardId,
2299 : timeline_id: TimelineId,
2300 : lsn: Lsn,
2301 : ctx: &RequestContext,
2302 : ) -> Result<(), QueryError>
2303 : where
2304 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
2305 : {
2306 : let timeline = self
2307 : .timeline_handles
2308 : .as_mut()
2309 : .unwrap()
2310 : .get(
2311 : tenant_shard_id.tenant_id,
2312 : timeline_id,
2313 : ShardSelector::Known(tenant_shard_id.to_index()),
2314 : )
2315 : .await?;
2316 : set_tracing_field_shard_id(&timeline);
2317 :
2318 : let lease = timeline
2319 : .renew_lsn_lease(lsn, timeline.get_lsn_lease_length(), ctx)
2320 0 : .inspect_err(|e| {
2321 0 : warn!("{e}");
2322 0 : })
2323 : .ok();
2324 0 : let valid_until_str = lease.map(|l| {
2325 0 : l.valid_until
2326 0 : .duration_since(SystemTime::UNIX_EPOCH)
2327 0 : .expect("valid_until is earlier than UNIX_EPOCH")
2328 0 : .as_millis()
2329 0 : .to_string()
2330 0 : });
2331 :
2332 : info!(
2333 : "acquired lease for {} until {}",
2334 : lsn,
2335 : valid_until_str.as_deref().unwrap_or("<unknown>")
2336 : );
2337 :
2338 0 : let bytes = valid_until_str.as_ref().map(|x| x.as_bytes());
2339 :
2340 : pgb.write_message_noflush(&BeMessage::RowDescription(&[RowDescriptor::text_col(
2341 : b"valid_until",
2342 : )]))?
2343 : .write_message_noflush(&BeMessage::DataRow(&[bytes]))?;
2344 :
2345 : Ok(())
2346 : }
2347 :
2348 : #[instrument(skip_all, fields(shard_id))]
2349 : async fn handle_get_rel_exists_request(
2350 : timeline: &Timeline,
2351 : req: &PagestreamExistsRequest,
2352 : ctx: &RequestContext,
2353 : ) -> Result<PagestreamExistsResponse, PageStreamError> {
2354 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
2355 : let lsn = Self::wait_or_get_last_lsn(
2356 : timeline,
2357 : req.hdr.request_lsn,
2358 : req.hdr.not_modified_since,
2359 : &latest_gc_cutoff_lsn,
2360 : ctx,
2361 : )
2362 : .await?;
2363 :
2364 : let exists = timeline
2365 : .get_rel_exists(
2366 : req.rel,
2367 : Version::LsnRange(LsnRange {
2368 : effective_lsn: lsn,
2369 : request_lsn: req.hdr.request_lsn,
2370 : }),
2371 : ctx,
2372 : )
2373 : .await?;
2374 :
2375 : Ok(PagestreamExistsResponse { req: *req, exists })
2376 : }
2377 :
2378 : /// If `allow_missing` is true, returns None instead of Err on missing relations. Otherwise,
2379 : /// never returns None. It is only supported by the gRPC protocol, so we pass it separately to
2380 : /// avoid changing the libpq protocol types.
2381 : #[instrument(skip_all, fields(shard_id))]
2382 : async fn handle_get_nblocks_request(
2383 : timeline: &Timeline,
2384 : req: &PagestreamNblocksRequest,
2385 : allow_missing: bool,
2386 : ctx: &RequestContext,
2387 : ) -> Result<Option<PagestreamNblocksResponse>, PageStreamError> {
2388 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
2389 : let lsn = Self::wait_or_get_last_lsn(
2390 : timeline,
2391 : req.hdr.request_lsn,
2392 : req.hdr.not_modified_since,
2393 : &latest_gc_cutoff_lsn,
2394 : ctx,
2395 : )
2396 : .await?;
2397 :
2398 : let n_blocks = timeline
2399 : .get_rel_size_in_reldir(
2400 : req.rel,
2401 : Version::LsnRange(LsnRange {
2402 : effective_lsn: lsn,
2403 : request_lsn: req.hdr.request_lsn,
2404 : }),
2405 : None,
2406 : allow_missing,
2407 : ctx,
2408 : )
2409 : .await?;
2410 : let Some(n_blocks) = n_blocks else {
2411 : return Ok(None);
2412 : };
2413 :
2414 : Ok(Some(PagestreamNblocksResponse {
2415 : req: *req,
2416 : n_blocks,
2417 : }))
2418 : }
2419 :
2420 : #[instrument(skip_all, fields(shard_id))]
2421 : async fn handle_db_size_request(
2422 : timeline: &Timeline,
2423 : req: &PagestreamDbSizeRequest,
2424 : ctx: &RequestContext,
2425 : ) -> Result<PagestreamDbSizeResponse, PageStreamError> {
2426 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
2427 : let lsn = Self::wait_or_get_last_lsn(
2428 : timeline,
2429 : req.hdr.request_lsn,
2430 : req.hdr.not_modified_since,
2431 : &latest_gc_cutoff_lsn,
2432 : ctx,
2433 : )
2434 : .await?;
2435 :
2436 : let total_blocks = timeline
2437 : .get_db_size(
2438 : DEFAULTTABLESPACE_OID,
2439 : req.dbnode,
2440 : Version::LsnRange(LsnRange {
2441 : effective_lsn: lsn,
2442 : request_lsn: req.hdr.request_lsn,
2443 : }),
2444 : ctx,
2445 : )
2446 : .await?;
2447 : let db_size = total_blocks as i64 * BLCKSZ as i64;
2448 :
2449 : Ok(PagestreamDbSizeResponse { req: *req, db_size })
2450 : }
2451 :
2452 : #[instrument(skip_all)]
2453 : async fn handle_get_page_at_lsn_request_batched(
2454 : timeline: &Timeline,
2455 : requests: SmallVec<[BatchedGetPageRequest; 1]>,
2456 : io_concurrency: IoConcurrency,
2457 : batch_break_reason: GetPageBatchBreakReason,
2458 : ctx: &RequestContext,
2459 : ) -> Vec<Result<(PagestreamBeMessage, SmgrOpTimer, RequestContext), BatchedPageStreamError>>
2460 : {
2461 : debug_assert_current_span_has_tenant_and_timeline_id();
2462 :
2463 : timeline
2464 : .query_metrics
2465 : .observe_getpage_batch_start(requests.len(), batch_break_reason);
2466 :
2467 : // If a page trace is running, submit an event for this request.
2468 : if let Some(page_trace) = timeline.page_trace.load().as_ref() {
2469 : let time = SystemTime::now();
2470 : for batch in &requests {
2471 : let key = rel_block_to_key(batch.req.rel, batch.req.blkno).to_compact();
2472 : // Ignore error (trace buffer may be full or tracer may have disconnected).
2473 : _ = page_trace.try_send(PageTraceEvent {
2474 : key,
2475 : effective_lsn: batch.lsn_range.effective_lsn,
2476 : time,
2477 : });
2478 : }
2479 : }
2480 :
2481 : // If any request in the batch needs to wait for LSN, then do so now.
2482 : let mut perf_instrument = false;
2483 : let max_effective_lsn = requests
2484 : .iter()
2485 0 : .map(|req| {
2486 0 : if req.ctx.has_perf_span() {
2487 0 : perf_instrument = true;
2488 0 : }
2489 :
2490 0 : req.lsn_range.effective_lsn
2491 0 : })
2492 : .max()
2493 : .expect("batch is never empty");
2494 :
2495 : let ctx = match perf_instrument {
2496 : true => RequestContextBuilder::from(ctx)
2497 0 : .root_perf_span(|| {
2498 0 : info_span!(
2499 : target: PERF_TRACE_TARGET,
2500 : "GET_VECTORED",
2501 : tenant_id = %timeline.tenant_shard_id.tenant_id,
2502 : timeline_id = %timeline.timeline_id,
2503 0 : shard = %timeline.tenant_shard_id.shard_slug(),
2504 : %max_effective_lsn
2505 : )
2506 0 : })
2507 : .attached_child(),
2508 : false => ctx.attached_child(),
2509 : };
2510 :
2511 : let last_record_lsn = timeline.get_last_record_lsn();
2512 : if max_effective_lsn > last_record_lsn {
2513 : if let Err(e) = timeline
2514 : .wait_lsn(
2515 : max_effective_lsn,
2516 : crate::tenant::timeline::WaitLsnWaiter::PageService,
2517 : timeline::WaitLsnTimeout::Default,
2518 : &ctx,
2519 : )
2520 0 : .maybe_perf_instrument(&ctx, |current_perf_span| {
2521 0 : info_span!(
2522 : target: PERF_TRACE_TARGET,
2523 0 : parent: current_perf_span,
2524 : "WAIT_LSN",
2525 : )
2526 0 : })
2527 : .await
2528 : {
2529 0 : return Vec::from_iter(requests.into_iter().map(|req| {
2530 0 : Err(BatchedPageStreamError {
2531 0 : err: PageStreamError::from(e.clone()),
2532 0 : req: req.req.hdr,
2533 0 : })
2534 0 : }));
2535 : }
2536 : }
2537 :
2538 : let results = timeline
2539 : .get_rel_page_at_lsn_batched(
2540 0 : requests.iter().map(|p| {
2541 0 : (
2542 0 : &p.req.rel,
2543 0 : &p.req.blkno,
2544 0 : p.lsn_range,
2545 0 : p.ctx.attached_child(),
2546 0 : )
2547 0 : }),
2548 : io_concurrency,
2549 : &ctx,
2550 : )
2551 : .await;
2552 : assert_eq!(results.len(), requests.len());
2553 :
2554 : // TODO: avoid creating the new Vec here
2555 : Vec::from_iter(
2556 : requests
2557 : .into_iter()
2558 : .zip(results.into_iter())
2559 0 : .map(|(req, res)| {
2560 0 : res.map(|page| {
2561 0 : (
2562 0 : PagestreamBeMessage::GetPage(
2563 0 : pagestream_api::PagestreamGetPageResponse { req: req.req, page },
2564 0 : ),
2565 0 : req.timer,
2566 0 : req.ctx,
2567 0 : )
2568 0 : })
2569 0 : .map_err(|e| BatchedPageStreamError {
2570 0 : err: PageStreamError::from(e),
2571 0 : req: req.req.hdr,
2572 0 : })
2573 0 : }),
2574 : )
2575 : }
2576 :
2577 : #[instrument(skip_all, fields(shard_id))]
2578 : async fn handle_get_slru_segment_request(
2579 : timeline: &Timeline,
2580 : req: &PagestreamGetSlruSegmentRequest,
2581 : ctx: &RequestContext,
2582 : ) -> Result<PagestreamGetSlruSegmentResponse, PageStreamError> {
2583 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
2584 : let lsn = Self::wait_or_get_last_lsn(
2585 : timeline,
2586 : req.hdr.request_lsn,
2587 : req.hdr.not_modified_since,
2588 : &latest_gc_cutoff_lsn,
2589 : ctx,
2590 : )
2591 : .await?;
2592 :
2593 : let kind = SlruKind::from_repr(req.kind)
2594 : .ok_or(PageStreamError::BadRequest("invalid SLRU kind".into()))?;
2595 : let segment = timeline.get_slru_segment(kind, req.segno, lsn, ctx).await?;
2596 :
2597 : Ok(PagestreamGetSlruSegmentResponse { req: *req, segment })
2598 : }
2599 :
2600 : // NB: this impl mimics what we do for batched getpage requests.
2601 : #[cfg(feature = "testing")]
2602 : #[instrument(skip_all, fields(shard_id))]
2603 : async fn handle_test_request_batch(
2604 : timeline: &Timeline,
2605 : requests: Vec<BatchedTestRequest>,
2606 : _ctx: &RequestContext,
2607 : ) -> Vec<Result<(PagestreamBeMessage, SmgrOpTimer, RequestContext), BatchedPageStreamError>>
2608 : {
2609 : // real requests would do something with the timeline
2610 : let mut results = Vec::with_capacity(requests.len());
2611 : for _req in requests.iter() {
2612 : tokio::task::yield_now().await;
2613 :
2614 : results.push({
2615 : if timeline.cancel.is_cancelled() {
2616 : Err(PageReconstructError::Cancelled)
2617 : } else {
2618 : Ok(())
2619 : }
2620 : });
2621 : }
2622 :
2623 : // TODO: avoid creating the new Vec here
2624 : Vec::from_iter(
2625 : requests
2626 : .into_iter()
2627 : .zip(results.into_iter())
2628 0 : .map(|(req, res)| {
2629 0 : res.map(|()| {
2630 0 : (
2631 0 : PagestreamBeMessage::Test(pagestream_api::PagestreamTestResponse {
2632 0 : req: req.req.clone(),
2633 0 : }),
2634 0 : req.timer,
2635 0 : RequestContext::new(
2636 0 : TaskKind::PageRequestHandler,
2637 0 : DownloadBehavior::Warn,
2638 0 : ),
2639 0 : )
2640 0 : })
2641 0 : .map_err(|e| BatchedPageStreamError {
2642 0 : err: PageStreamError::from(e),
2643 0 : req: req.req.hdr,
2644 0 : })
2645 0 : }),
2646 : )
2647 : }
2648 :
2649 : /// Note on "fullbackup":
2650 : /// Full basebackups should only be used for debugging purposes.
2651 : /// Originally, it was introduced to enable breaking storage format changes,
2652 : /// but that is not applicable anymore.
2653 : ///
2654 : /// # Coding Discipline
2655 : ///
2656 : /// Coding discipline within this function: all interaction with the `pgb` connection
2657 : /// needs to be sensitive to connection shutdown, currently signalled via [`Self::cancel`].
2658 : /// This is so that we can shutdown page_service quickly.
2659 : ///
2660 : /// TODO: wrap the pgb that we pass to the basebackup handler so that it's sensitive
2661 : /// to connection cancellation.
2662 : #[allow(clippy::too_many_arguments)]
2663 : #[instrument(skip_all, fields(shard_id, ?lsn, ?prev_lsn, %full_backup))]
2664 : async fn handle_basebackup_request<IO>(
2665 : &mut self,
2666 : pgb: &mut PostgresBackend<IO>,
2667 : tenant_id: TenantId,
2668 : timeline_id: TimelineId,
2669 : lsn: Option<Lsn>,
2670 : prev_lsn: Option<Lsn>,
2671 : full_backup: bool,
2672 : gzip: bool,
2673 : replica: bool,
2674 : ctx: &RequestContext,
2675 : ) -> Result<(), QueryError>
2676 : where
2677 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
2678 : {
2679 : let started = std::time::Instant::now();
2680 :
2681 : let timeline = self
2682 : .timeline_handles
2683 : .as_mut()
2684 : .unwrap()
2685 : .get(tenant_id, timeline_id, ShardSelector::Zero)
2686 : .await?;
2687 : set_tracing_field_shard_id(&timeline);
2688 : let ctx = ctx.with_scope_timeline(&timeline);
2689 :
2690 : if timeline.is_archived() == Some(true) {
2691 : tracing::info!(
2692 : "timeline {tenant_id}/{timeline_id} is archived, but got basebackup request for it."
2693 : );
2694 : return Err(QueryError::NotFound("timeline is archived".into()));
2695 : }
2696 :
2697 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
2698 : if let Some(lsn) = lsn {
2699 : // Backup was requested at a particular LSN. Wait for it to arrive.
2700 : info!("waiting for {}", lsn);
2701 : timeline
2702 : .wait_lsn(
2703 : lsn,
2704 : crate::tenant::timeline::WaitLsnWaiter::PageService,
2705 : crate::tenant::timeline::WaitLsnTimeout::Default,
2706 : &ctx,
2707 : )
2708 : .await?;
2709 : timeline
2710 : .check_lsn_is_in_scope(lsn, &latest_gc_cutoff_lsn)
2711 : .context("invalid basebackup lsn")?;
2712 : }
2713 :
2714 : let lsn_awaited_after = started.elapsed();
2715 :
2716 : // switch client to COPYOUT
2717 : pgb.write_message_noflush(&BeMessage::CopyOutResponse)
2718 : .map_err(QueryError::Disconnected)?;
2719 : self.flush_cancellable(pgb, &self.cancel).await?;
2720 :
2721 : let mut from_cache = false;
2722 :
2723 : // Send a tarball of the latest layer on the timeline. Compress if not
2724 : // fullbackup. TODO Compress in that case too (tests need to be updated)
2725 : if full_backup {
2726 : let mut writer = pgb.copyout_writer();
2727 : basebackup::send_basebackup_tarball(
2728 : &mut writer,
2729 : &timeline,
2730 : lsn,
2731 : prev_lsn,
2732 : full_backup,
2733 : replica,
2734 : None,
2735 : &ctx,
2736 : )
2737 : .await?;
2738 : } else {
2739 : let mut writer = BufWriter::new(pgb.copyout_writer());
2740 :
2741 : let cached = timeline
2742 : .get_cached_basebackup_if_enabled(lsn, prev_lsn, full_backup, replica, gzip)
2743 : .await;
2744 :
2745 : if let Some(mut cached) = cached {
2746 : from_cache = true;
2747 : tokio::io::copy(&mut cached, &mut writer)
2748 : .await
2749 0 : .map_err(|err| {
2750 0 : BasebackupError::Client(err, "handle_basebackup_request,cached,copy")
2751 0 : })?;
2752 : } else {
2753 : basebackup::send_basebackup_tarball(
2754 : &mut writer,
2755 : &timeline,
2756 : lsn,
2757 : prev_lsn,
2758 : full_backup,
2759 : replica,
2760 : // NB: using fast compression because it's on the critical path for compute
2761 : // startup. For an empty database, we get <100KB with this method. The
2762 : // Level::Best compression method gives us <20KB, but maybe we should add
2763 : // basebackup caching on compute shutdown first.
2764 : gzip.then_some(async_compression::Level::Fastest),
2765 : &ctx,
2766 : )
2767 : .await?;
2768 : }
2769 : writer
2770 : .flush()
2771 : .await
2772 0 : .map_err(|err| BasebackupError::Client(err, "handle_basebackup_request,flush"))?;
2773 : }
2774 :
2775 : pgb.write_message_noflush(&BeMessage::CopyDone)
2776 : .map_err(QueryError::Disconnected)?;
2777 : self.flush_cancellable(pgb, &timeline.cancel).await?;
2778 :
2779 : let basebackup_after = started
2780 : .elapsed()
2781 : .checked_sub(lsn_awaited_after)
2782 : .unwrap_or(Duration::ZERO);
2783 :
2784 : info!(
2785 : lsn_await_millis = lsn_awaited_after.as_millis(),
2786 : basebackup_millis = basebackup_after.as_millis(),
2787 : %from_cache,
2788 : "basebackup complete"
2789 : );
2790 :
2791 : Ok(())
2792 : }
2793 :
2794 : // when accessing management api supply None as an argument
2795 : // when using to authorize tenant pass corresponding tenant id
2796 0 : fn check_permission(&self, tenant_id: Option<TenantId>) -> Result<(), QueryError> {
2797 0 : if self.auth.is_none() {
2798 : // auth is set to Trust, nothing to check so just return ok
2799 0 : return Ok(());
2800 0 : }
2801 : // auth is some, just checked above, when auth is some
2802 : // then claims are always present because of checks during connection init
2803 : // so this expect won't trigger
2804 0 : let claims = self
2805 0 : .claims
2806 0 : .as_ref()
2807 0 : .expect("claims presence already checked");
2808 0 : check_permission(claims, tenant_id).map_err(|e| QueryError::Unauthorized(e.0))
2809 0 : }
2810 : }
2811 :
2812 : /// `basebackup tenant timeline [lsn] [--gzip] [--replica]`
2813 : #[derive(Debug, Clone, Eq, PartialEq)]
2814 : struct BaseBackupCmd {
2815 : tenant_id: TenantId,
2816 : timeline_id: TimelineId,
2817 : lsn: Option<Lsn>,
2818 : gzip: bool,
2819 : replica: bool,
2820 : }
2821 :
2822 : /// `fullbackup tenant timeline [lsn] [prev_lsn]`
2823 : #[derive(Debug, Clone, Eq, PartialEq)]
2824 : struct FullBackupCmd {
2825 : tenant_id: TenantId,
2826 : timeline_id: TimelineId,
2827 : lsn: Option<Lsn>,
2828 : prev_lsn: Option<Lsn>,
2829 : }
2830 :
2831 : /// `pagestream_v2 tenant timeline`
2832 : #[derive(Debug, Clone, Eq, PartialEq)]
2833 : struct PageStreamCmd {
2834 : tenant_id: TenantId,
2835 : timeline_id: TimelineId,
2836 : protocol_version: PagestreamProtocolVersion,
2837 : }
2838 :
2839 : /// `lease lsn tenant timeline lsn`
2840 : #[derive(Debug, Clone, Eq, PartialEq)]
2841 : struct LeaseLsnCmd {
2842 : tenant_shard_id: TenantShardId,
2843 : timeline_id: TimelineId,
2844 : lsn: Lsn,
2845 : }
2846 :
2847 : #[derive(Debug, Clone, Eq, PartialEq)]
2848 : enum PageServiceCmd {
2849 : Set,
2850 : PageStream(PageStreamCmd),
2851 : BaseBackup(BaseBackupCmd),
2852 : FullBackup(FullBackupCmd),
2853 : LeaseLsn(LeaseLsnCmd),
2854 : }
2855 :
2856 : impl PageStreamCmd {
2857 3 : fn parse(query: &str, protocol_version: PagestreamProtocolVersion) -> anyhow::Result<Self> {
2858 3 : let parameters = query.split_whitespace().collect_vec();
2859 3 : if parameters.len() != 2 {
2860 1 : bail!(
2861 1 : "invalid number of parameters for pagestream command: {}",
2862 : query
2863 : );
2864 2 : }
2865 2 : let tenant_id = TenantId::from_str(parameters[0])
2866 2 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
2867 1 : let timeline_id = TimelineId::from_str(parameters[1])
2868 1 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
2869 1 : Ok(Self {
2870 1 : tenant_id,
2871 1 : timeline_id,
2872 1 : protocol_version,
2873 1 : })
2874 3 : }
2875 : }
2876 :
2877 : impl FullBackupCmd {
2878 2 : fn parse(query: &str) -> anyhow::Result<Self> {
2879 2 : let parameters = query.split_whitespace().collect_vec();
2880 2 : if parameters.len() < 2 || parameters.len() > 4 {
2881 0 : bail!(
2882 0 : "invalid number of parameters for basebackup command: {}",
2883 : query
2884 : );
2885 2 : }
2886 2 : let tenant_id = TenantId::from_str(parameters[0])
2887 2 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
2888 2 : let timeline_id = TimelineId::from_str(parameters[1])
2889 2 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
2890 : // The caller is responsible for providing correct lsn and prev_lsn.
2891 2 : let lsn = if let Some(lsn_str) = parameters.get(2) {
2892 : Some(
2893 1 : Lsn::from_str(lsn_str)
2894 1 : .with_context(|| format!("Failed to parse Lsn from {lsn_str}"))?,
2895 : )
2896 : } else {
2897 1 : None
2898 : };
2899 2 : let prev_lsn = if let Some(prev_lsn_str) = parameters.get(3) {
2900 : Some(
2901 1 : Lsn::from_str(prev_lsn_str)
2902 1 : .with_context(|| format!("Failed to parse Lsn from {prev_lsn_str}"))?,
2903 : )
2904 : } else {
2905 1 : None
2906 : };
2907 2 : Ok(Self {
2908 2 : tenant_id,
2909 2 : timeline_id,
2910 2 : lsn,
2911 2 : prev_lsn,
2912 2 : })
2913 2 : }
2914 : }
2915 :
2916 : impl BaseBackupCmd {
2917 9 : fn parse(query: &str) -> anyhow::Result<Self> {
2918 9 : let parameters = query.split_whitespace().collect_vec();
2919 9 : if parameters.len() < 2 {
2920 0 : bail!(
2921 0 : "invalid number of parameters for basebackup command: {}",
2922 : query
2923 : );
2924 9 : }
2925 9 : let tenant_id = TenantId::from_str(parameters[0])
2926 9 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
2927 9 : let timeline_id = TimelineId::from_str(parameters[1])
2928 9 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
2929 : let lsn;
2930 : let flags_parse_from;
2931 9 : if let Some(maybe_lsn) = parameters.get(2) {
2932 8 : if *maybe_lsn == "latest" {
2933 1 : lsn = None;
2934 1 : flags_parse_from = 3;
2935 7 : } else if maybe_lsn.starts_with("--") {
2936 5 : lsn = None;
2937 5 : flags_parse_from = 2;
2938 5 : } else {
2939 : lsn = Some(
2940 2 : Lsn::from_str(maybe_lsn)
2941 2 : .with_context(|| format!("Failed to parse lsn from {maybe_lsn}"))?,
2942 : );
2943 2 : flags_parse_from = 3;
2944 : }
2945 1 : } else {
2946 1 : lsn = None;
2947 1 : flags_parse_from = 2;
2948 1 : }
2949 :
2950 9 : let mut gzip = false;
2951 9 : let mut replica = false;
2952 :
2953 11 : for ¶m in ¶meters[flags_parse_from..] {
2954 11 : match param {
2955 11 : "--gzip" => {
2956 7 : if gzip {
2957 1 : bail!("duplicate parameter for basebackup command: {param}")
2958 6 : }
2959 6 : gzip = true
2960 : }
2961 4 : "--replica" => {
2962 2 : if replica {
2963 0 : bail!("duplicate parameter for basebackup command: {param}")
2964 2 : }
2965 2 : replica = true
2966 : }
2967 2 : _ => bail!("invalid parameter for basebackup command: {param}"),
2968 : }
2969 : }
2970 6 : Ok(Self {
2971 6 : tenant_id,
2972 6 : timeline_id,
2973 6 : lsn,
2974 6 : gzip,
2975 6 : replica,
2976 6 : })
2977 9 : }
2978 : }
2979 :
2980 : impl LeaseLsnCmd {
2981 2 : fn parse(query: &str) -> anyhow::Result<Self> {
2982 2 : let parameters = query.split_whitespace().collect_vec();
2983 2 : if parameters.len() != 3 {
2984 0 : bail!(
2985 0 : "invalid number of parameters for lease lsn command: {}",
2986 : query
2987 : );
2988 2 : }
2989 2 : let tenant_shard_id = TenantShardId::from_str(parameters[0])
2990 2 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
2991 2 : let timeline_id = TimelineId::from_str(parameters[1])
2992 2 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
2993 2 : let lsn = Lsn::from_str(parameters[2])
2994 2 : .with_context(|| format!("Failed to parse lsn from {}", parameters[2]))?;
2995 2 : Ok(Self {
2996 2 : tenant_shard_id,
2997 2 : timeline_id,
2998 2 : lsn,
2999 2 : })
3000 2 : }
3001 : }
3002 :
3003 : impl PageServiceCmd {
3004 21 : fn parse(query: &str) -> anyhow::Result<Self> {
3005 21 : let query = query.trim();
3006 21 : let Some((cmd, other)) = query.split_once(' ') else {
3007 2 : bail!("cannot parse query: {query}")
3008 : };
3009 19 : match cmd.to_ascii_lowercase().as_str() {
3010 19 : "pagestream_v2" => Ok(Self::PageStream(PageStreamCmd::parse(
3011 3 : other,
3012 3 : PagestreamProtocolVersion::V2,
3013 2 : )?)),
3014 16 : "pagestream_v3" => Ok(Self::PageStream(PageStreamCmd::parse(
3015 0 : other,
3016 0 : PagestreamProtocolVersion::V3,
3017 0 : )?)),
3018 16 : "basebackup" => Ok(Self::BaseBackup(BaseBackupCmd::parse(other)?)),
3019 7 : "fullbackup" => Ok(Self::FullBackup(FullBackupCmd::parse(other)?)),
3020 5 : "lease" => {
3021 3 : let Some((cmd2, other)) = other.split_once(' ') else {
3022 0 : bail!("invalid lease command: {cmd}");
3023 : };
3024 3 : let cmd2 = cmd2.to_ascii_lowercase();
3025 3 : if cmd2 == "lsn" {
3026 2 : Ok(Self::LeaseLsn(LeaseLsnCmd::parse(other)?))
3027 : } else {
3028 1 : bail!("invalid lease command: {cmd}");
3029 : }
3030 : }
3031 2 : "set" => Ok(Self::Set),
3032 0 : _ => Err(anyhow::anyhow!("unsupported command {cmd} in {query}")),
3033 : }
3034 21 : }
3035 : }
3036 :
3037 : /// Parse the startup options from the postgres wire protocol startup packet.
3038 : ///
3039 : /// It takes a sequence of `-c option=X` or `-coption=X`. It parses the options string
3040 : /// by best effort and returns all the options parsed (key-value pairs) and a bool indicating
3041 : /// whether all options are successfully parsed. There could be duplicates in the options
3042 : /// if the caller passed such parameters.
3043 7 : fn parse_options(options: &str) -> (Vec<(String, String)>, bool) {
3044 7 : let mut parsing_config = false;
3045 7 : let mut has_error = false;
3046 7 : let mut config = Vec::new();
3047 16 : for item in options.split_whitespace() {
3048 16 : if item == "-c" {
3049 9 : if !parsing_config {
3050 8 : parsing_config = true;
3051 8 : } else {
3052 : // "-c" followed with another "-c"
3053 1 : tracing::warn!("failed to parse the startup options: {options}");
3054 1 : has_error = true;
3055 1 : break;
3056 : }
3057 7 : } else if item.starts_with("-c") || parsing_config {
3058 7 : let Some((mut key, value)) = item.split_once('=') else {
3059 : // "-c" followed with an invalid option
3060 1 : tracing::warn!("failed to parse the startup options: {options}");
3061 1 : has_error = true;
3062 1 : break;
3063 : };
3064 6 : if !parsing_config {
3065 : // Parse "-coptions=X"
3066 1 : let Some(stripped_key) = key.strip_prefix("-c") else {
3067 0 : tracing::warn!("failed to parse the startup options: {options}");
3068 0 : has_error = true;
3069 0 : break;
3070 : };
3071 1 : key = stripped_key;
3072 5 : }
3073 6 : config.push((key.to_string(), value.to_string()));
3074 6 : parsing_config = false;
3075 : } else {
3076 0 : tracing::warn!("failed to parse the startup options: {options}");
3077 0 : has_error = true;
3078 0 : break;
3079 : }
3080 : }
3081 7 : if parsing_config {
3082 : // "-c" without the option
3083 3 : tracing::warn!("failed to parse the startup options: {options}");
3084 3 : has_error = true;
3085 4 : }
3086 7 : (config, has_error)
3087 7 : }
3088 :
3089 : impl<IO> postgres_backend::Handler<IO> for PageServerHandler
3090 : where
3091 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
3092 : {
3093 0 : fn check_auth_jwt(
3094 0 : &mut self,
3095 0 : _pgb: &mut PostgresBackend<IO>,
3096 0 : jwt_response: &[u8],
3097 0 : ) -> Result<(), QueryError> {
3098 : // this unwrap is never triggered, because check_auth_jwt only called when auth_type is NeonJWT
3099 : // which requires auth to be present
3100 0 : let data: TokenData<Claims> = self
3101 0 : .auth
3102 0 : .as_ref()
3103 0 : .unwrap()
3104 0 : .decode(str::from_utf8(jwt_response).context("jwt response is not UTF-8")?)
3105 0 : .map_err(|e| QueryError::Unauthorized(e.0))?;
3106 :
3107 0 : if matches!(data.claims.scope, Scope::Tenant) && data.claims.tenant_id.is_none() {
3108 0 : return Err(QueryError::Unauthorized(
3109 0 : "jwt token scope is Tenant, but tenant id is missing".into(),
3110 0 : ));
3111 0 : }
3112 :
3113 0 : debug!(
3114 0 : "jwt scope check succeeded for scope: {:#?} by tenant id: {:?}",
3115 : data.claims.scope, data.claims.tenant_id,
3116 : );
3117 :
3118 0 : self.claims = Some(data.claims);
3119 0 : Ok(())
3120 0 : }
3121 :
3122 0 : fn startup(
3123 0 : &mut self,
3124 0 : _pgb: &mut PostgresBackend<IO>,
3125 0 : sm: &FeStartupPacket,
3126 0 : ) -> Result<(), QueryError> {
3127 0 : fail::fail_point!("ps::connection-start::startup-packet");
3128 :
3129 0 : if let FeStartupPacket::StartupMessage { params, .. } = sm {
3130 0 : if let Some(app_name) = params.get("application_name") {
3131 0 : self.perf_span_fields.application_name = Some(app_name.to_string());
3132 0 : Span::current().record("application_name", field::display(app_name));
3133 0 : }
3134 0 : if let Some(options) = params.get("options") {
3135 0 : let (config, _) = parse_options(options);
3136 0 : for (key, value) in config {
3137 0 : if key == "neon.compute_mode" {
3138 0 : self.perf_span_fields.compute_mode = Some(value.clone());
3139 0 : Span::current().record("compute_mode", field::display(value));
3140 0 : }
3141 : }
3142 0 : }
3143 0 : };
3144 :
3145 0 : Ok(())
3146 0 : }
3147 :
3148 : #[instrument(skip_all, fields(tenant_id, timeline_id))]
3149 : async fn process_query(
3150 : &mut self,
3151 : pgb: &mut PostgresBackend<IO>,
3152 : query_string: &str,
3153 : ) -> Result<(), QueryError> {
3154 0 : fail::fail_point!("simulated-bad-compute-connection", |_| {
3155 0 : info!("Hit failpoint for bad connection");
3156 0 : Err(QueryError::SimulatedConnectionError)
3157 0 : });
3158 :
3159 : fail::fail_point!("ps::connection-start::process-query");
3160 :
3161 : let ctx = self.connection_ctx.attached_child();
3162 : debug!("process query {query_string}");
3163 : let query = PageServiceCmd::parse(query_string)?;
3164 : match query {
3165 : PageServiceCmd::PageStream(PageStreamCmd {
3166 : tenant_id,
3167 : timeline_id,
3168 : protocol_version,
3169 : }) => {
3170 : tracing::Span::current()
3171 : .record("tenant_id", field::display(tenant_id))
3172 : .record("timeline_id", field::display(timeline_id));
3173 :
3174 : self.check_permission(Some(tenant_id))?;
3175 : let command_kind = match protocol_version {
3176 : PagestreamProtocolVersion::V2 => ComputeCommandKind::PageStreamV2,
3177 : PagestreamProtocolVersion::V3 => ComputeCommandKind::PageStreamV3,
3178 : };
3179 : COMPUTE_COMMANDS_COUNTERS.for_command(command_kind).inc();
3180 :
3181 : self.handle_pagerequests(pgb, tenant_id, timeline_id, protocol_version, ctx)
3182 : .await?;
3183 : }
3184 : PageServiceCmd::BaseBackup(BaseBackupCmd {
3185 : tenant_id,
3186 : timeline_id,
3187 : lsn,
3188 : gzip,
3189 : replica,
3190 : }) => {
3191 : tracing::Span::current()
3192 : .record("tenant_id", field::display(tenant_id))
3193 : .record("timeline_id", field::display(timeline_id));
3194 :
3195 : self.check_permission(Some(tenant_id))?;
3196 :
3197 : COMPUTE_COMMANDS_COUNTERS
3198 : .for_command(ComputeCommandKind::Basebackup)
3199 : .inc();
3200 : let metric_recording = metrics::BASEBACKUP_QUERY_TIME.start_recording();
3201 0 : let res = async {
3202 0 : self.handle_basebackup_request(
3203 0 : pgb,
3204 0 : tenant_id,
3205 0 : timeline_id,
3206 0 : lsn,
3207 0 : None,
3208 0 : false,
3209 0 : gzip,
3210 0 : replica,
3211 0 : &ctx,
3212 0 : )
3213 0 : .await?;
3214 0 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
3215 0 : Result::<(), QueryError>::Ok(())
3216 0 : }
3217 : .await;
3218 : metric_recording.observe(&res);
3219 : res?;
3220 : }
3221 : // same as basebackup, but result includes relational data as well
3222 : PageServiceCmd::FullBackup(FullBackupCmd {
3223 : tenant_id,
3224 : timeline_id,
3225 : lsn,
3226 : prev_lsn,
3227 : }) => {
3228 : tracing::Span::current()
3229 : .record("tenant_id", field::display(tenant_id))
3230 : .record("timeline_id", field::display(timeline_id));
3231 :
3232 : self.check_permission(Some(tenant_id))?;
3233 :
3234 : COMPUTE_COMMANDS_COUNTERS
3235 : .for_command(ComputeCommandKind::Fullbackup)
3236 : .inc();
3237 :
3238 : // Check that the timeline exists
3239 : self.handle_basebackup_request(
3240 : pgb,
3241 : tenant_id,
3242 : timeline_id,
3243 : lsn,
3244 : prev_lsn,
3245 : true,
3246 : false,
3247 : false,
3248 : &ctx,
3249 : )
3250 : .await?;
3251 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
3252 : }
3253 : PageServiceCmd::Set => {
3254 : // important because psycopg2 executes "SET datestyle TO 'ISO'"
3255 : // on connect
3256 : // TODO: allow setting options, i.e., application_name/compute_mode via SET commands
3257 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
3258 : }
3259 : PageServiceCmd::LeaseLsn(LeaseLsnCmd {
3260 : tenant_shard_id,
3261 : timeline_id,
3262 : lsn,
3263 : }) => {
3264 : tracing::Span::current()
3265 : .record("tenant_id", field::display(tenant_shard_id))
3266 : .record("timeline_id", field::display(timeline_id));
3267 :
3268 : self.check_permission(Some(tenant_shard_id.tenant_id))?;
3269 :
3270 : COMPUTE_COMMANDS_COUNTERS
3271 : .for_command(ComputeCommandKind::LeaseLsn)
3272 : .inc();
3273 :
3274 : match self
3275 : .handle_make_lsn_lease(pgb, tenant_shard_id, timeline_id, lsn, &ctx)
3276 : .await
3277 : {
3278 : Ok(()) => {
3279 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?
3280 : }
3281 : Err(e) => {
3282 : error!("error obtaining lsn lease for {lsn}: {e:?}");
3283 : pgb.write_message_noflush(&BeMessage::ErrorResponse(
3284 : &e.to_string(),
3285 : Some(e.pg_error_code()),
3286 : ))?
3287 : }
3288 : };
3289 : }
3290 : }
3291 :
3292 : Ok(())
3293 : }
3294 : }
3295 :
3296 : /// Serves the page service over gRPC. Dispatches to PageServerHandler for request processing.
3297 : ///
3298 : /// TODO: rename to PageServiceHandler when libpq impl is removed.
3299 : pub struct GrpcPageServiceHandler {
3300 : tenant_manager: Arc<TenantManager>,
3301 : ctx: RequestContext,
3302 :
3303 : /// Cancelled to shut down the server. Tonic will shut down in response to this, but wait for
3304 : /// in-flight requests to complete. Any tasks we spawn ourselves must respect this token.
3305 : cancel: CancellationToken,
3306 :
3307 : /// Any tasks we spawn ourselves should clone this gate guard, so that we can wait for them to
3308 : /// complete during shutdown. Request handlers implicitly hold this guard already.
3309 : gate_guard: GateGuard,
3310 :
3311 : /// `get_vectored` concurrency setting.
3312 : get_vectored_concurrent_io: GetVectoredConcurrentIo,
3313 : }
3314 :
3315 : impl GrpcPageServiceHandler {
3316 : /// Spawns a gRPC server for the page service.
3317 : ///
3318 : /// Returns a `CancellableTask` handle that can be used to shut down the server. It waits for
3319 : /// any in-flight requests and tasks to complete first.
3320 : ///
3321 : /// TODO: this doesn't support TLS. We need TLS reloading via ReloadingCertificateResolver, so we
3322 : /// need to reimplement the TCP+TLS accept loop ourselves.
3323 0 : pub fn spawn(
3324 0 : tenant_manager: Arc<TenantManager>,
3325 0 : auth: Option<Arc<SwappableJwtAuth>>,
3326 0 : perf_trace_dispatch: Option<Dispatch>,
3327 0 : get_vectored_concurrent_io: GetVectoredConcurrentIo,
3328 0 : listener: std::net::TcpListener,
3329 0 : ) -> anyhow::Result<CancellableTask> {
3330 : // Set up a cancellation token for shutting down the server, and a gate to wait for all
3331 : // requests and spawned tasks to complete.
3332 0 : let cancel = CancellationToken::new();
3333 0 : let gate = Gate::default();
3334 :
3335 0 : let ctx = RequestContextBuilder::new(TaskKind::PageRequestHandler)
3336 0 : .download_behavior(DownloadBehavior::Download)
3337 0 : .perf_span_dispatch(perf_trace_dispatch)
3338 0 : .detached_child();
3339 :
3340 : // Set up the TCP socket. We take a preconfigured TcpListener to bind the
3341 : // port early during startup.
3342 0 : let incoming = {
3343 0 : let _runtime = COMPUTE_REQUEST_RUNTIME.enter(); // required by TcpListener::from_std
3344 0 : listener.set_nonblocking(true)?;
3345 0 : tonic::transport::server::TcpIncoming::from(tokio::net::TcpListener::from_std(
3346 0 : listener,
3347 0 : )?)
3348 0 : .with_nodelay(Some(GRPC_TCP_NODELAY))
3349 0 : .with_keepalive(Some(GRPC_TCP_KEEPALIVE_TIME))
3350 : };
3351 :
3352 : // Set up the gRPC server.
3353 : //
3354 : // TODO: consider tuning window sizes.
3355 0 : let mut server = tonic::transport::Server::builder()
3356 0 : .http2_keepalive_interval(Some(GRPC_HTTP2_KEEPALIVE_INTERVAL))
3357 0 : .http2_keepalive_timeout(Some(GRPC_HTTP2_KEEPALIVE_TIMEOUT))
3358 0 : .max_concurrent_streams(Some(GRPC_MAX_CONCURRENT_STREAMS));
3359 :
3360 : // Main page service stack. Uses a mix of Tonic interceptors and Tower layers:
3361 : //
3362 : // * Interceptors: can inspect and modify the gRPC request. Sync code only, runs before service.
3363 : //
3364 : // * Layers: allow async code, can run code after the service response. However, only has access
3365 : // to the raw HTTP request/response, not the gRPC types.
3366 0 : let page_service_handler = GrpcPageServiceHandler {
3367 0 : tenant_manager,
3368 0 : ctx,
3369 0 : cancel: cancel.clone(),
3370 0 : gate_guard: gate.enter().expect("gate was just created"),
3371 0 : get_vectored_concurrent_io,
3372 0 : };
3373 :
3374 0 : let observability_layer = ObservabilityLayer;
3375 0 : let mut tenant_interceptor = TenantMetadataInterceptor;
3376 0 : let mut auth_interceptor = TenantAuthInterceptor::new(auth);
3377 :
3378 0 : let page_service = tower::ServiceBuilder::new()
3379 : // Create tracing span and record request start time.
3380 0 : .layer(observability_layer)
3381 : // Intercept gRPC requests.
3382 0 : .layer(tonic::service::InterceptorLayer::new(move |mut req| {
3383 : // Extract tenant metadata.
3384 0 : req = tenant_interceptor.call(req)?;
3385 : // Authenticate tenant JWT token.
3386 0 : req = auth_interceptor.call(req)?;
3387 0 : Ok(req)
3388 0 : }))
3389 : // Run the page service.
3390 0 : .service(
3391 0 : proto::PageServiceServer::new(page_service_handler)
3392 : // Support both gzip and zstd compression. The client decides what to use.
3393 0 : .accept_compressed(tonic::codec::CompressionEncoding::Gzip)
3394 0 : .accept_compressed(tonic::codec::CompressionEncoding::Zstd)
3395 0 : .send_compressed(tonic::codec::CompressionEncoding::Gzip)
3396 0 : .send_compressed(tonic::codec::CompressionEncoding::Zstd),
3397 : );
3398 0 : let server = server.add_service(page_service);
3399 :
3400 : // Reflection service for use with e.g. grpcurl.
3401 0 : let reflection_service = tonic_reflection::server::Builder::configure()
3402 0 : .register_encoded_file_descriptor_set(proto::FILE_DESCRIPTOR_SET)
3403 0 : .build_v1()?;
3404 0 : let server = server.add_service(reflection_service);
3405 :
3406 : // Spawn server task. It runs until the cancellation token fires and in-flight requests and
3407 : // tasks complete. The `CancellableTask` will wait for the task's join handle, which
3408 : // implicitly waits for the gate to close.
3409 0 : let task_cancel = cancel.clone();
3410 0 : let task = COMPUTE_REQUEST_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
3411 : "grpc pageservice listener",
3412 0 : async move {
3413 0 : server
3414 0 : .serve_with_incoming_shutdown(incoming, task_cancel.cancelled())
3415 0 : .await?;
3416 : // Server exited cleanly. All requests should have completed by now. Wait for any
3417 : // spawned tasks to complete as well (e.g. IoConcurrency sidecars) via the gate.
3418 0 : gate.close().await;
3419 0 : anyhow::Ok(())
3420 0 : },
3421 : ));
3422 :
3423 0 : Ok(CancellableTask { task, cancel })
3424 0 : }
3425 :
3426 : /// Errors if the request is executed on a non-zero shard. Only shard 0 has a complete view of
3427 : /// relations and their sizes, as well as SLRU segments and similar data.
3428 : #[allow(clippy::result_large_err)]
3429 0 : fn ensure_shard_zero(timeline: &Handle<TenantManagerTypes>) -> Result<(), tonic::Status> {
3430 0 : match timeline.get_shard_index().shard_number.0 {
3431 0 : 0 => Ok(()),
3432 0 : shard => Err(tonic::Status::invalid_argument(format!(
3433 0 : "request must execute on shard zero (is shard {shard})",
3434 0 : ))),
3435 : }
3436 0 : }
3437 :
3438 : /// Generates a PagestreamRequest header from a ReadLsn and request ID.
3439 0 : fn make_hdr(
3440 0 : read_lsn: page_api::ReadLsn,
3441 0 : req_id: Option<page_api::RequestID>,
3442 0 : ) -> PagestreamRequest {
3443 : PagestreamRequest {
3444 0 : reqid: req_id.map(|r| r.id).unwrap_or_default(),
3445 0 : request_lsn: read_lsn.request_lsn,
3446 0 : not_modified_since: read_lsn
3447 0 : .not_modified_since_lsn
3448 0 : .unwrap_or(read_lsn.request_lsn),
3449 : }
3450 0 : }
3451 :
3452 : /// Acquires a timeline handle for the given request.
3453 : ///
3454 : /// TODO: during shard splits, the compute may still be sending requests to the parent shard
3455 : /// until the entire split is committed and the compute is notified. Consider installing a
3456 : /// temporary shard router from the parent to the children while the split is in progress.
3457 : ///
3458 : /// TODO: consider moving this to a middleware layer; all requests need it. Needs to manage
3459 : /// the TimelineHandles lifecycle.
3460 : ///
3461 : /// TODO: untangle acquisition from TenantManagerWrapper::resolve() and Cache::get(), to avoid
3462 : /// the unnecessary overhead.
3463 0 : async fn get_request_timeline(
3464 0 : &self,
3465 0 : req: &tonic::Request<impl Any>,
3466 0 : ) -> Result<Handle<TenantManagerTypes>, GetActiveTimelineError> {
3467 0 : let ttid = *extract::<TenantTimelineId>(req);
3468 0 : let shard_index = *extract::<ShardIndex>(req);
3469 0 : let shard_selector = ShardSelector::Known(shard_index);
3470 :
3471 0 : TimelineHandles::new(self.tenant_manager.clone())
3472 0 : .get(ttid.tenant_id, ttid.timeline_id, shard_selector)
3473 0 : .await
3474 0 : }
3475 :
3476 : /// Starts a SmgrOpTimer at received_at, throttles the request, and records execution start.
3477 : /// Only errors if the timeline is shutting down.
3478 : ///
3479 : /// TODO: move timer construction to ObservabilityLayer (see TODO there).
3480 : /// TODO: decouple rate limiting (middleware?), and return SlowDown errors instead.
3481 0 : async fn record_op_start_and_throttle(
3482 0 : timeline: &Handle<TenantManagerTypes>,
3483 0 : op: metrics::SmgrQueryType,
3484 0 : received_at: Instant,
3485 0 : ) -> Result<SmgrOpTimer, tonic::Status> {
3486 0 : let mut timer = PageServerHandler::record_op_start_and_throttle(timeline, op, received_at)
3487 0 : .await
3488 0 : .map_err(|err| match err {
3489 : // record_op_start_and_throttle() only returns Shutdown.
3490 0 : QueryError::Shutdown => tonic::Status::unavailable(format!("{err}")),
3491 0 : err => tonic::Status::internal(format!("unexpected error: {err}")),
3492 0 : })?;
3493 0 : timer.observe_execution_start(Instant::now());
3494 0 : Ok(timer)
3495 0 : }
3496 :
3497 : /// Processes a GetPage batch request, via the GetPages bidirectional streaming RPC.
3498 : ///
3499 : /// NB: errors returned from here are intercepted in get_pages(), and may be converted to a
3500 : /// GetPageResponse with an appropriate status code to avoid terminating the stream.
3501 : ///
3502 : /// TODO: get_vectored() currently enforces a batch limit of 32. Postgres will typically send
3503 : /// batches up to effective_io_concurrency = 100. Either we have to accept large batches, or
3504 : /// split them up in the client or server.
3505 : #[instrument(skip_all, fields(req_id, rel, blkno, blks, req_lsn, mod_lsn))]
3506 : async fn get_page(
3507 : ctx: &RequestContext,
3508 : timeline: &WeakHandle<TenantManagerTypes>,
3509 : req: proto::GetPageRequest,
3510 : io_concurrency: IoConcurrency,
3511 : ) -> Result<proto::GetPageResponse, tonic::Status> {
3512 : let received_at = Instant::now();
3513 : let timeline = timeline.upgrade()?;
3514 : let ctx = ctx.with_scope_page_service_pagestream(&timeline);
3515 :
3516 : // Validate the request, decorate the span, and convert it to a Pagestream request.
3517 : let req = page_api::GetPageRequest::try_from(req)?;
3518 :
3519 : span_record!(
3520 : req_id = %req.request_id,
3521 : rel = %req.rel,
3522 : blkno = %req.block_numbers[0],
3523 : blks = %req.block_numbers.len(),
3524 : lsn = %req.read_lsn,
3525 : );
3526 :
3527 : for &blkno in &req.block_numbers {
3528 : let shard = timeline.get_shard_identity();
3529 : let key = rel_block_to_key(req.rel, blkno);
3530 : if !shard.is_key_local(&key) {
3531 : return Err(tonic::Status::invalid_argument(format!(
3532 : "block {blkno} of relation {} requested on wrong shard {} (is on {})",
3533 : req.rel,
3534 : timeline.get_shard_index(),
3535 : ShardIndex::new(shard.get_shard_number(&key), shard.count),
3536 : )));
3537 : }
3538 : }
3539 :
3540 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn(); // hold guard
3541 : let effective_lsn = PageServerHandler::effective_request_lsn(
3542 : &timeline,
3543 : timeline.get_last_record_lsn(),
3544 : req.read_lsn.request_lsn,
3545 : req.read_lsn
3546 : .not_modified_since_lsn
3547 : .unwrap_or(req.read_lsn.request_lsn),
3548 : &latest_gc_cutoff_lsn,
3549 : )?;
3550 :
3551 : let mut batch = SmallVec::with_capacity(req.block_numbers.len());
3552 : for blkno in req.block_numbers {
3553 : // TODO: this creates one timer per page and throttles it. We should have a timer for
3554 : // the entire batch, and throttle only the batch, but this is equivalent to what
3555 : // PageServerHandler does already so we keep it for now.
3556 : let timer = Self::record_op_start_and_throttle(
3557 : &timeline,
3558 : metrics::SmgrQueryType::GetPageAtLsn,
3559 : received_at,
3560 : )
3561 : .await?;
3562 :
3563 : batch.push(BatchedGetPageRequest {
3564 : req: PagestreamGetPageRequest {
3565 : hdr: Self::make_hdr(req.read_lsn, Some(req.request_id)),
3566 : rel: req.rel,
3567 : blkno,
3568 : },
3569 : lsn_range: LsnRange {
3570 : effective_lsn,
3571 : request_lsn: req.read_lsn.request_lsn,
3572 : },
3573 : timer,
3574 : ctx: ctx.attached_child(),
3575 : batch_wait_ctx: None, // TODO: add tracing
3576 : });
3577 : }
3578 :
3579 : // TODO: this does a relation size query for every page in the batch. Since this batch is
3580 : // all for one relation, we could do this only once. However, this is not the case for the
3581 : // libpq implementation.
3582 : let results = PageServerHandler::handle_get_page_at_lsn_request_batched(
3583 : &timeline,
3584 : batch,
3585 : io_concurrency,
3586 : GetPageBatchBreakReason::BatchFull, // TODO: not relevant for gRPC batches
3587 : &ctx,
3588 : )
3589 : .await;
3590 :
3591 : let mut resp = page_api::GetPageResponse {
3592 : request_id: req.request_id,
3593 : status_code: page_api::GetPageStatusCode::Ok,
3594 : reason: None,
3595 : rel: req.rel,
3596 : pages: Vec::with_capacity(results.len()),
3597 : };
3598 :
3599 : for result in results {
3600 : match result {
3601 : Ok((PagestreamBeMessage::GetPage(r), _, _)) => resp.pages.push(page_api::Page {
3602 : block_number: r.req.blkno,
3603 : image: r.page,
3604 : }),
3605 : Ok((resp, _, _)) => {
3606 : return Err(tonic::Status::internal(format!(
3607 : "unexpected response: {resp:?}"
3608 : )));
3609 : }
3610 : Err(err) => return Err(err.err.into()),
3611 : };
3612 : }
3613 :
3614 : Ok(resp.into())
3615 : }
3616 : }
3617 :
3618 : /// Implements the gRPC page service.
3619 : ///
3620 : /// On client disconnect (e.g. timeout or client shutdown), Tonic will drop the request handler
3621 : /// futures, so the read path must be cancellation-safe. On server shutdown, Tonic will wait for
3622 : /// in-flight requests to complete.
3623 : ///
3624 : /// TODO: when the libpq impl is removed, remove the Pagestream types and inline the handler code.
3625 : #[tonic::async_trait]
3626 : impl proto::PageService for GrpcPageServiceHandler {
3627 : type GetBaseBackupStream = Pin<
3628 : Box<dyn Stream<Item = Result<proto::GetBaseBackupResponseChunk, tonic::Status>> + Send>,
3629 : >;
3630 :
3631 : type GetPagesStream =
3632 : Pin<Box<dyn Stream<Item = Result<proto::GetPageResponse, tonic::Status>> + Send>>;
3633 :
3634 : #[instrument(skip_all, fields(lsn))]
3635 : async fn get_base_backup(
3636 : &self,
3637 : req: tonic::Request<proto::GetBaseBackupRequest>,
3638 0 : ) -> Result<tonic::Response<Self::GetBaseBackupStream>, tonic::Status> {
3639 : // Send chunks of 256 KB to avoid large memory allocations. pagebench basebackup shows this
3640 : // to be the sweet spot where throughput is saturated.
3641 : const CHUNK_SIZE: usize = 256 * 1024;
3642 :
3643 0 : let timeline = self.get_request_timeline(&req).await?;
3644 0 : let ctx = self.ctx.with_scope_timeline(&timeline);
3645 :
3646 : // Validate the request and decorate the span.
3647 0 : Self::ensure_shard_zero(&timeline)?;
3648 0 : if timeline.is_archived() == Some(true) {
3649 0 : return Err(tonic::Status::failed_precondition("timeline is archived"));
3650 0 : }
3651 0 : let req: page_api::GetBaseBackupRequest = req.into_inner().try_into()?;
3652 :
3653 0 : span_record!(lsn=?req.lsn);
3654 :
3655 : // Wait for the LSN to arrive, if given.
3656 0 : if let Some(lsn) = req.lsn {
3657 0 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
3658 0 : timeline
3659 0 : .wait_lsn(
3660 0 : lsn,
3661 0 : WaitLsnWaiter::PageService,
3662 0 : WaitLsnTimeout::Default,
3663 0 : &ctx,
3664 0 : )
3665 0 : .await?;
3666 0 : timeline
3667 0 : .check_lsn_is_in_scope(lsn, &latest_gc_cutoff_lsn)
3668 0 : .map_err(|err| {
3669 0 : tonic::Status::invalid_argument(format!("invalid basebackup LSN: {err}"))
3670 0 : })?;
3671 0 : }
3672 :
3673 : // Spawn a task to run the basebackup.
3674 0 : let span = Span::current();
3675 0 : let gate_guard = self
3676 0 : .gate_guard
3677 0 : .try_clone()
3678 0 : .map_err(|_| tonic::Status::unavailable("shutting down"))?;
3679 0 : let (mut simplex_read, mut simplex_write) = tokio::io::simplex(CHUNK_SIZE);
3680 0 : let jh = tokio::spawn(async move {
3681 0 : let _gate_guard = gate_guard; // keep gate open until task completes
3682 :
3683 0 : let gzip_level = match req.compression {
3684 0 : page_api::BaseBackupCompression::None => None,
3685 : // NB: using fast compression because it's on the critical path for compute
3686 : // startup. For an empty database, we get <100KB with this method. The
3687 : // Level::Best compression method gives us <20KB, but maybe we should add
3688 : // basebackup caching on compute shutdown first.
3689 0 : page_api::BaseBackupCompression::Gzip => Some(async_compression::Level::Fastest),
3690 : };
3691 :
3692 : // Check for a cached basebackup.
3693 0 : let cached = timeline
3694 0 : .get_cached_basebackup_if_enabled(
3695 0 : req.lsn,
3696 0 : None,
3697 0 : req.full,
3698 0 : req.replica,
3699 0 : gzip_level.is_some(),
3700 0 : )
3701 0 : .await;
3702 :
3703 0 : let result = if let Some(mut cached) = cached {
3704 : // If we have a cached basebackup, send it.
3705 0 : tokio::io::copy(&mut cached, &mut simplex_write)
3706 0 : .await
3707 0 : .map(|_| ())
3708 0 : .map_err(|err| BasebackupError::Client(err, "cached,copy"))
3709 : } else {
3710 0 : basebackup::send_basebackup_tarball(
3711 0 : &mut simplex_write,
3712 0 : &timeline,
3713 0 : req.lsn,
3714 0 : None,
3715 0 : req.full,
3716 0 : req.replica,
3717 0 : gzip_level,
3718 0 : &ctx,
3719 0 : )
3720 0 : .instrument(span) // propagate request span
3721 0 : .await
3722 : };
3723 0 : simplex_write
3724 0 : .shutdown()
3725 0 : .await
3726 0 : .map_err(|err| BasebackupError::Client(err, "simplex_write"))?;
3727 0 : result
3728 0 : });
3729 :
3730 : // Emit chunks of size CHUNK_SIZE.
3731 0 : let chunks = async_stream::try_stream! {
3732 : loop {
3733 : let mut chunk = BytesMut::with_capacity(CHUNK_SIZE).limit(CHUNK_SIZE);
3734 : loop {
3735 0 : let n = simplex_read.read_buf(&mut chunk).await.map_err(|err| {
3736 0 : tonic::Status::internal(format!("failed to read basebackup chunk: {err}"))
3737 0 : })?;
3738 : if n == 0 {
3739 : break; // full chunk or closed stream
3740 : }
3741 : }
3742 : let chunk = chunk.into_inner().freeze();
3743 : if chunk.is_empty() {
3744 : break;
3745 : }
3746 : yield proto::GetBaseBackupResponseChunk::from(chunk);
3747 : }
3748 : // Wait for the basebackup task to exit and check for errors.
3749 0 : jh.await.map_err(|err| {
3750 0 : tonic::Status::internal(format!("basebackup failed: {err}"))
3751 0 : })??;
3752 : };
3753 :
3754 0 : Ok(tonic::Response::new(Box::pin(chunks)))
3755 0 : }
3756 :
3757 : #[instrument(skip_all, fields(db_oid, lsn))]
3758 : async fn get_db_size(
3759 : &self,
3760 : req: tonic::Request<proto::GetDbSizeRequest>,
3761 0 : ) -> Result<tonic::Response<proto::GetDbSizeResponse>, tonic::Status> {
3762 0 : let received_at = extract::<ReceivedAt>(&req).0;
3763 0 : let timeline = self.get_request_timeline(&req).await?;
3764 0 : let ctx = self.ctx.with_scope_page_service_pagestream(&timeline);
3765 :
3766 : // Validate the request, decorate the span, and convert it to a Pagestream request.
3767 0 : Self::ensure_shard_zero(&timeline)?;
3768 0 : let req: page_api::GetDbSizeRequest = req.into_inner().try_into()?;
3769 :
3770 0 : span_record!(db_oid=%req.db_oid, lsn=%req.read_lsn);
3771 :
3772 0 : let req = PagestreamDbSizeRequest {
3773 0 : hdr: Self::make_hdr(req.read_lsn, None),
3774 0 : dbnode: req.db_oid,
3775 0 : };
3776 :
3777 : // Execute the request and convert the response.
3778 0 : let _timer = Self::record_op_start_and_throttle(
3779 0 : &timeline,
3780 0 : metrics::SmgrQueryType::GetDbSize,
3781 0 : received_at,
3782 0 : )
3783 0 : .await?;
3784 :
3785 0 : let resp = PageServerHandler::handle_db_size_request(&timeline, &req, &ctx).await?;
3786 0 : let resp = resp.db_size as page_api::GetDbSizeResponse;
3787 0 : Ok(tonic::Response::new(resp.into()))
3788 0 : }
3789 :
3790 : // NB: don't instrument this, instrument each streamed request.
3791 0 : async fn get_pages(
3792 : &self,
3793 : req: tonic::Request<tonic::Streaming<proto::GetPageRequest>>,
3794 0 : ) -> Result<tonic::Response<Self::GetPagesStream>, tonic::Status> {
3795 : // Extract the timeline from the request and check that it exists.
3796 0 : let ttid = *extract::<TenantTimelineId>(&req);
3797 0 : let shard_index = *extract::<ShardIndex>(&req);
3798 0 : let shard_selector = ShardSelector::Known(shard_index);
3799 :
3800 0 : let mut handles = TimelineHandles::new(self.tenant_manager.clone());
3801 0 : handles
3802 0 : .get(ttid.tenant_id, ttid.timeline_id, shard_selector)
3803 0 : .await?;
3804 :
3805 : // Spawn an IoConcurrency sidecar, if enabled.
3806 0 : let gate_guard = self
3807 0 : .gate_guard
3808 0 : .try_clone()
3809 0 : .map_err(|_| tonic::Status::unavailable("shutting down"))?;
3810 0 : let io_concurrency =
3811 0 : IoConcurrency::spawn_from_conf(self.get_vectored_concurrent_io, gate_guard);
3812 :
3813 : // Construct the GetPageRequest stream handler.
3814 0 : let span = Span::current();
3815 0 : let ctx = self.ctx.attached_child();
3816 0 : let cancel = self.cancel.clone();
3817 0 : let mut reqs = req.into_inner();
3818 :
3819 0 : let resps = async_stream::try_stream! {
3820 : let timeline = handles
3821 : .get(ttid.tenant_id, ttid.timeline_id, shard_selector)
3822 : .await?
3823 : .downgrade();
3824 : loop {
3825 : // NB: Tonic considers the entire stream to be an in-flight request and will wait
3826 : // for it to complete before shutting down. React to cancellation between requests.
3827 : let req = tokio::select! {
3828 : biased;
3829 : _ = cancel.cancelled() => Err(tonic::Status::unavailable("shutting down")),
3830 :
3831 : result = reqs.message() => match result {
3832 : Ok(Some(req)) => Ok(req),
3833 : Ok(None) => break, // client closed the stream
3834 : Err(err) => Err(err),
3835 : },
3836 : }?;
3837 : let req_id = req.request_id.map(page_api::RequestID::from).unwrap_or_default();
3838 : let result = Self::get_page(&ctx, &timeline, req, io_concurrency.clone())
3839 : .instrument(span.clone()) // propagate request span
3840 : .await;
3841 : yield match result {
3842 : Ok(resp) => resp,
3843 : // Convert per-request errors to GetPageResponses as appropriate, or terminate
3844 : // the stream with a tonic::Status. Log the error regardless, since
3845 : // ObservabilityLayer can't automatically log stream errors.
3846 : Err(status) => {
3847 : // TODO: it would be nice if we could propagate the get_page() fields here.
3848 0 : span.in_scope(|| {
3849 0 : warn!("request failed with {:?}: {}", status.code(), status.message());
3850 0 : });
3851 : page_api::GetPageResponse::try_from_status(status, req_id)?.into()
3852 : }
3853 : }
3854 : }
3855 : };
3856 :
3857 0 : Ok(tonic::Response::new(Box::pin(resps)))
3858 0 : }
3859 :
3860 : #[instrument(skip_all, fields(rel, lsn, allow_missing))]
3861 : async fn get_rel_size(
3862 : &self,
3863 : req: tonic::Request<proto::GetRelSizeRequest>,
3864 0 : ) -> Result<tonic::Response<proto::GetRelSizeResponse>, tonic::Status> {
3865 0 : let received_at = extract::<ReceivedAt>(&req).0;
3866 0 : let timeline = self.get_request_timeline(&req).await?;
3867 0 : let ctx = self.ctx.with_scope_page_service_pagestream(&timeline);
3868 :
3869 : // Validate the request, decorate the span, and convert it to a Pagestream request.
3870 0 : Self::ensure_shard_zero(&timeline)?;
3871 0 : let req: page_api::GetRelSizeRequest = req.into_inner().try_into()?;
3872 0 : let allow_missing = req.allow_missing;
3873 :
3874 0 : span_record!(rel=%req.rel, lsn=%req.read_lsn, allow_missing=%req.allow_missing);
3875 :
3876 0 : let req = PagestreamNblocksRequest {
3877 0 : hdr: Self::make_hdr(req.read_lsn, None),
3878 0 : rel: req.rel,
3879 0 : };
3880 :
3881 : // Execute the request and convert the response.
3882 0 : let _timer = Self::record_op_start_and_throttle(
3883 0 : &timeline,
3884 0 : metrics::SmgrQueryType::GetRelSize,
3885 0 : received_at,
3886 0 : )
3887 0 : .await?;
3888 :
3889 0 : let resp =
3890 0 : PageServerHandler::handle_get_nblocks_request(&timeline, &req, allow_missing, &ctx)
3891 0 : .await?;
3892 0 : let resp: page_api::GetRelSizeResponse = resp.map(|resp| resp.n_blocks);
3893 :
3894 0 : Ok(tonic::Response::new(resp.into()))
3895 0 : }
3896 :
3897 : #[instrument(skip_all, fields(kind, segno, lsn))]
3898 : async fn get_slru_segment(
3899 : &self,
3900 : req: tonic::Request<proto::GetSlruSegmentRequest>,
3901 0 : ) -> Result<tonic::Response<proto::GetSlruSegmentResponse>, tonic::Status> {
3902 0 : let received_at = extract::<ReceivedAt>(&req).0;
3903 0 : let timeline = self.get_request_timeline(&req).await?;
3904 0 : let ctx = self.ctx.with_scope_page_service_pagestream(&timeline);
3905 :
3906 : // Validate the request, decorate the span, and convert it to a Pagestream request.
3907 0 : Self::ensure_shard_zero(&timeline)?;
3908 0 : let req: page_api::GetSlruSegmentRequest = req.into_inner().try_into()?;
3909 :
3910 0 : span_record!(kind=%req.kind, segno=%req.segno, lsn=%req.read_lsn);
3911 :
3912 0 : let req = PagestreamGetSlruSegmentRequest {
3913 0 : hdr: Self::make_hdr(req.read_lsn, None),
3914 0 : kind: req.kind as u8,
3915 0 : segno: req.segno,
3916 0 : };
3917 :
3918 : // Execute the request and convert the response.
3919 0 : let _timer = Self::record_op_start_and_throttle(
3920 0 : &timeline,
3921 0 : metrics::SmgrQueryType::GetSlruSegment,
3922 0 : received_at,
3923 0 : )
3924 0 : .await?;
3925 :
3926 0 : let resp =
3927 0 : PageServerHandler::handle_get_slru_segment_request(&timeline, &req, &ctx).await?;
3928 0 : let resp: page_api::GetSlruSegmentResponse = resp.segment;
3929 0 : Ok(tonic::Response::new(resp.into()))
3930 0 : }
3931 :
3932 : #[instrument(skip_all, fields(lsn))]
3933 : async fn lease_lsn(
3934 : &self,
3935 : req: tonic::Request<proto::LeaseLsnRequest>,
3936 0 : ) -> Result<tonic::Response<proto::LeaseLsnResponse>, tonic::Status> {
3937 0 : let timeline = self.get_request_timeline(&req).await?;
3938 0 : let ctx = self.ctx.with_scope_timeline(&timeline);
3939 :
3940 : // Validate and convert the request, and decorate the span.
3941 0 : let req: page_api::LeaseLsnRequest = req.into_inner().try_into()?;
3942 :
3943 0 : span_record!(lsn=%req.lsn);
3944 :
3945 : // Attempt to acquire a lease. Return FailedPrecondition if the lease could not be granted.
3946 0 : let lease_length = timeline.get_lsn_lease_length();
3947 0 : let expires = match timeline.renew_lsn_lease(req.lsn, lease_length, &ctx) {
3948 0 : Ok(lease) => lease.valid_until,
3949 0 : Err(err) => return Err(tonic::Status::failed_precondition(format!("{err}"))),
3950 : };
3951 :
3952 : // TODO: is this spammy? Move it compute-side?
3953 0 : info!(
3954 0 : "acquired lease for {} until {}",
3955 : req.lsn,
3956 0 : chrono::DateTime::<Utc>::from(expires).to_rfc3339()
3957 : );
3958 :
3959 0 : Ok(tonic::Response::new(expires.into()))
3960 0 : }
3961 : }
3962 :
3963 : /// gRPC middleware layer that handles observability concerns:
3964 : ///
3965 : /// * Creates and enters a tracing span.
3966 : /// * Records the request start time as a ReceivedAt request extension.
3967 : ///
3968 : /// TODO: add perf tracing.
3969 : /// TODO: add timing and metrics.
3970 : /// TODO: add logging.
3971 : #[derive(Clone)]
3972 : struct ObservabilityLayer;
3973 :
3974 : impl<S: tonic::server::NamedService> tower::Layer<S> for ObservabilityLayer {
3975 : type Service = ObservabilityLayerService<S>;
3976 :
3977 0 : fn layer(&self, inner: S) -> Self::Service {
3978 0 : Self::Service { inner }
3979 0 : }
3980 : }
3981 :
3982 : #[derive(Clone)]
3983 : struct ObservabilityLayerService<S> {
3984 : inner: S,
3985 : }
3986 :
3987 : #[derive(Clone, Copy)]
3988 : struct ReceivedAt(Instant);
3989 :
3990 : impl<S: tonic::server::NamedService> tonic::server::NamedService for ObservabilityLayerService<S> {
3991 : const NAME: &'static str = S::NAME; // propagate inner service name
3992 : }
3993 :
3994 : impl<S, Req, Resp> tower::Service<http::Request<Req>> for ObservabilityLayerService<S>
3995 : where
3996 : S: tower::Service<http::Request<Req>, Response = http::Response<Resp>> + Send,
3997 : S::Future: Send + 'static,
3998 : {
3999 : type Response = S::Response;
4000 : type Error = S::Error;
4001 : type Future = BoxFuture<'static, Result<Self::Response, Self::Error>>;
4002 :
4003 0 : fn call(&mut self, mut req: http::Request<Req>) -> Self::Future {
4004 : // Record the request start time as a request extension.
4005 : //
4006 : // TODO: we should start a timer here instead, but it currently requires a timeline handle
4007 : // and SmgrQueryType, which we don't have yet. Refactor it to provide it later.
4008 0 : req.extensions_mut().insert(ReceivedAt(Instant::now()));
4009 :
4010 : // Extract the peer address and gRPC method.
4011 0 : let peer = req
4012 0 : .extensions()
4013 0 : .get::<TcpConnectInfo>()
4014 0 : .and_then(|info| info.remote_addr())
4015 0 : .map(|addr| addr.to_string())
4016 0 : .unwrap_or_default();
4017 :
4018 0 : let method = req
4019 0 : .uri()
4020 0 : .path()
4021 0 : .split('/')
4022 0 : .nth(2)
4023 0 : .unwrap_or(req.uri().path())
4024 0 : .to_string();
4025 :
4026 : // Create a basic tracing span.
4027 : //
4028 : // Enter the span for the current thread and instrument the future. It is not sufficient to
4029 : // only instrument the future, since it only takes effect after the future is returned and
4030 : // polled, not when the inner service is called below (e.g. during interceptor execution).
4031 0 : let span = info_span!(
4032 : "grpc:pageservice",
4033 : // These will be populated by TenantMetadataInterceptor.
4034 : tenant_id = field::Empty,
4035 : timeline_id = field::Empty,
4036 : shard_id = field::Empty,
4037 : // NB: empty fields must be listed first above. Otherwise, the field names will be
4038 : // clobbered when the empty fields are populated. They will be output last regardless.
4039 : %peer,
4040 : %method,
4041 : );
4042 0 : let _guard = span.enter();
4043 :
4044 : // Construct a future for calling the inner service, but don't await it. This avoids having
4045 : // to clone the inner service into the future below.
4046 0 : let call = self.inner.call(req);
4047 :
4048 0 : async move {
4049 : // Await the inner service call.
4050 0 : let result = call.await;
4051 :
4052 : // Log gRPC error statuses. This won't include request info from handler spans, but it
4053 : // will catch all errors (even those emitted before handler spans are constructed). Only
4054 : // unary request errors are logged here, not streaming response errors.
4055 0 : if let Ok(ref resp) = result
4056 0 : && let Some(status) = tonic::Status::from_header_map(resp.headers())
4057 0 : && status.code() != tonic::Code::Ok
4058 : {
4059 : // TODO: it would be nice if we could propagate the handler span's request fields
4060 : // here. This could e.g. be done by attaching the request fields to
4061 : // tonic::Status::metadata via a proc macro.
4062 0 : warn!(
4063 0 : "request failed with {:?}: {}",
4064 0 : status.code(),
4065 0 : status.message()
4066 : );
4067 0 : }
4068 :
4069 0 : result
4070 0 : }
4071 0 : .instrument(span.clone())
4072 0 : .boxed()
4073 0 : }
4074 :
4075 0 : fn poll_ready(&mut self, cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
4076 0 : self.inner.poll_ready(cx)
4077 0 : }
4078 : }
4079 :
4080 : /// gRPC interceptor that decodes tenant metadata and stores it as request extensions of type
4081 : /// TenantTimelineId and ShardIndex.
4082 : #[derive(Clone)]
4083 : struct TenantMetadataInterceptor;
4084 :
4085 : impl tonic::service::Interceptor for TenantMetadataInterceptor {
4086 0 : fn call(&mut self, mut req: tonic::Request<()>) -> Result<tonic::Request<()>, tonic::Status> {
4087 : // Decode the tenant ID.
4088 0 : let tenant_id = req
4089 0 : .metadata()
4090 0 : .get("neon-tenant-id")
4091 0 : .ok_or_else(|| tonic::Status::invalid_argument("missing neon-tenant-id"))?
4092 0 : .to_str()
4093 0 : .map_err(|_| tonic::Status::invalid_argument("invalid neon-tenant-id"))?;
4094 0 : let tenant_id = TenantId::from_str(tenant_id)
4095 0 : .map_err(|_| tonic::Status::invalid_argument("invalid neon-tenant-id"))?;
4096 :
4097 : // Decode the timeline ID.
4098 0 : let timeline_id = req
4099 0 : .metadata()
4100 0 : .get("neon-timeline-id")
4101 0 : .ok_or_else(|| tonic::Status::invalid_argument("missing neon-timeline-id"))?
4102 0 : .to_str()
4103 0 : .map_err(|_| tonic::Status::invalid_argument("invalid neon-timeline-id"))?;
4104 0 : let timeline_id = TimelineId::from_str(timeline_id)
4105 0 : .map_err(|_| tonic::Status::invalid_argument("invalid neon-timeline-id"))?;
4106 :
4107 : // Decode the shard ID.
4108 0 : let shard_id = req
4109 0 : .metadata()
4110 0 : .get("neon-shard-id")
4111 0 : .ok_or_else(|| tonic::Status::invalid_argument("missing neon-shard-id"))?
4112 0 : .to_str()
4113 0 : .map_err(|_| tonic::Status::invalid_argument("invalid neon-shard-id"))?;
4114 0 : let shard_id = ShardIndex::from_str(shard_id)
4115 0 : .map_err(|_| tonic::Status::invalid_argument("invalid neon-shard-id"))?;
4116 :
4117 : // Stash them in the request.
4118 0 : let extensions = req.extensions_mut();
4119 0 : extensions.insert(TenantTimelineId::new(tenant_id, timeline_id));
4120 0 : extensions.insert(shard_id);
4121 :
4122 : // Decorate the tracing span.
4123 0 : span_record!(%tenant_id, %timeline_id, %shard_id);
4124 :
4125 0 : Ok(req)
4126 0 : }
4127 : }
4128 :
4129 : /// Authenticates gRPC page service requests.
4130 : #[derive(Clone)]
4131 : struct TenantAuthInterceptor {
4132 : auth: Option<Arc<SwappableJwtAuth>>,
4133 : }
4134 :
4135 : impl TenantAuthInterceptor {
4136 0 : fn new(auth: Option<Arc<SwappableJwtAuth>>) -> Self {
4137 0 : Self { auth }
4138 0 : }
4139 : }
4140 :
4141 : impl tonic::service::Interceptor for TenantAuthInterceptor {
4142 0 : fn call(&mut self, req: tonic::Request<()>) -> Result<tonic::Request<()>, tonic::Status> {
4143 : // Do nothing if auth is disabled.
4144 0 : let Some(auth) = self.auth.as_ref() else {
4145 0 : return Ok(req);
4146 : };
4147 :
4148 : // Fetch the tenant ID from the request extensions (set by TenantMetadataInterceptor).
4149 0 : let TenantTimelineId { tenant_id, .. } = *extract::<TenantTimelineId>(&req);
4150 :
4151 : // Fetch and decode the JWT token.
4152 0 : let jwt = req
4153 0 : .metadata()
4154 0 : .get("authorization")
4155 0 : .ok_or_else(|| tonic::Status::unauthenticated("no authorization header"))?
4156 0 : .to_str()
4157 0 : .map_err(|_| tonic::Status::invalid_argument("invalid authorization header"))?
4158 0 : .strip_prefix("Bearer ")
4159 0 : .ok_or_else(|| tonic::Status::invalid_argument("invalid authorization header"))?
4160 0 : .trim();
4161 0 : let jwtdata: TokenData<Claims> = auth
4162 0 : .decode(jwt)
4163 0 : .map_err(|err| tonic::Status::invalid_argument(format!("invalid JWT token: {err}")))?;
4164 0 : let claims = jwtdata.claims;
4165 :
4166 : // Check if the token is valid for this tenant.
4167 0 : check_permission(&claims, Some(tenant_id))
4168 0 : .map_err(|err| tonic::Status::permission_denied(err.to_string()))?;
4169 :
4170 : // TODO: consider stashing the claims in the request extensions, if needed.
4171 :
4172 0 : Ok(req)
4173 0 : }
4174 : }
4175 :
4176 : /// Extracts the given type from the request extensions, or panics if it is missing.
4177 0 : fn extract<T: Send + Sync + 'static>(req: &tonic::Request<impl Any>) -> &T {
4178 0 : extract_from(req.extensions())
4179 0 : }
4180 :
4181 : /// Extract the given type from the request extensions, or panics if it is missing. This variant
4182 : /// can extract both from a tonic::Request and http::Request.
4183 0 : fn extract_from<T: Send + Sync + 'static>(ext: &http::Extensions) -> &T {
4184 0 : let Some(value) = ext.get::<T>() else {
4185 0 : let name = std::any::type_name::<T>();
4186 0 : panic!("extension {name} should be set by middleware");
4187 : };
4188 0 : value
4189 0 : }
4190 :
4191 : #[derive(Debug, thiserror::Error)]
4192 : pub(crate) enum GetActiveTimelineError {
4193 : #[error(transparent)]
4194 : Tenant(GetActiveTenantError),
4195 : #[error(transparent)]
4196 : Timeline(#[from] GetTimelineError),
4197 : }
4198 :
4199 : impl From<GetActiveTimelineError> for QueryError {
4200 0 : fn from(e: GetActiveTimelineError) -> Self {
4201 0 : match e {
4202 0 : GetActiveTimelineError::Tenant(GetActiveTenantError::Cancelled) => QueryError::Shutdown,
4203 0 : GetActiveTimelineError::Tenant(e) => e.into(),
4204 0 : GetActiveTimelineError::Timeline(e) => QueryError::NotFound(format!("{e}").into()),
4205 : }
4206 0 : }
4207 : }
4208 :
4209 : impl From<GetActiveTimelineError> for tonic::Status {
4210 0 : fn from(err: GetActiveTimelineError) -> Self {
4211 0 : let message = err.to_string();
4212 0 : let code = match err {
4213 0 : GetActiveTimelineError::Tenant(err) => tonic::Status::from(err).code(),
4214 0 : GetActiveTimelineError::Timeline(err) => tonic::Status::from(err).code(),
4215 : };
4216 0 : tonic::Status::new(code, message)
4217 0 : }
4218 : }
4219 :
4220 : impl From<GetTimelineError> for tonic::Status {
4221 0 : fn from(err: GetTimelineError) -> Self {
4222 : use tonic::Code;
4223 0 : let code = match &err {
4224 0 : GetTimelineError::NotFound { .. } => Code::NotFound,
4225 0 : GetTimelineError::NotActive { .. } => Code::Unavailable,
4226 0 : GetTimelineError::ShuttingDown => Code::Unavailable,
4227 : };
4228 0 : tonic::Status::new(code, err.to_string())
4229 0 : }
4230 : }
4231 :
4232 : impl From<GetActiveTenantError> for QueryError {
4233 0 : fn from(e: GetActiveTenantError) -> Self {
4234 0 : match e {
4235 0 : GetActiveTenantError::WaitForActiveTimeout { .. } => QueryError::Disconnected(
4236 0 : ConnectionError::Io(io::Error::new(io::ErrorKind::TimedOut, e.to_string())),
4237 0 : ),
4238 : GetActiveTenantError::Cancelled
4239 : | GetActiveTenantError::WillNotBecomeActive(TenantState::Stopping { .. }) => {
4240 0 : QueryError::Shutdown
4241 : }
4242 0 : e @ GetActiveTenantError::NotFound(_) => QueryError::NotFound(format!("{e}").into()),
4243 0 : e => QueryError::Other(anyhow::anyhow!(e)),
4244 : }
4245 0 : }
4246 : }
4247 :
4248 : impl From<GetActiveTenantError> for tonic::Status {
4249 0 : fn from(err: GetActiveTenantError) -> Self {
4250 : use tonic::Code;
4251 0 : let code = match &err {
4252 0 : GetActiveTenantError::Broken(_) => Code::Internal,
4253 0 : GetActiveTenantError::Cancelled => Code::Unavailable,
4254 0 : GetActiveTenantError::NotFound(_) => Code::NotFound,
4255 0 : GetActiveTenantError::SwitchedTenant => Code::Unavailable,
4256 0 : GetActiveTenantError::WaitForActiveTimeout { .. } => Code::Unavailable,
4257 0 : GetActiveTenantError::WillNotBecomeActive(_) => Code::Unavailable,
4258 : };
4259 0 : tonic::Status::new(code, err.to_string())
4260 0 : }
4261 : }
4262 :
4263 : impl From<HandleUpgradeError> for QueryError {
4264 0 : fn from(e: HandleUpgradeError) -> Self {
4265 0 : match e {
4266 0 : HandleUpgradeError::ShutDown => QueryError::Shutdown,
4267 : }
4268 0 : }
4269 : }
4270 :
4271 : impl From<HandleUpgradeError> for tonic::Status {
4272 0 : fn from(err: HandleUpgradeError) -> Self {
4273 0 : match err {
4274 0 : HandleUpgradeError::ShutDown => tonic::Status::unavailable("timeline is shutting down"),
4275 : }
4276 0 : }
4277 : }
4278 :
4279 0 : fn set_tracing_field_shard_id(timeline: &Timeline) {
4280 0 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
4281 0 : tracing::Span::current().record(
4282 0 : "shard_id",
4283 0 : tracing::field::display(timeline.tenant_shard_id.shard_slug()),
4284 : );
4285 0 : debug_assert_current_span_has_tenant_and_timeline_id();
4286 0 : }
4287 :
4288 : struct WaitedForLsn(Lsn);
4289 : impl From<WaitedForLsn> for Lsn {
4290 0 : fn from(WaitedForLsn(lsn): WaitedForLsn) -> Self {
4291 0 : lsn
4292 0 : }
4293 : }
4294 :
4295 : #[cfg(test)]
4296 : mod tests {
4297 : use utils::shard::ShardCount;
4298 :
4299 : use super::*;
4300 :
4301 : #[test]
4302 1 : fn pageservice_cmd_parse() {
4303 1 : let tenant_id = TenantId::generate();
4304 1 : let timeline_id = TimelineId::generate();
4305 1 : let cmd =
4306 1 : PageServiceCmd::parse(&format!("pagestream_v2 {tenant_id} {timeline_id}")).unwrap();
4307 1 : assert_eq!(
4308 : cmd,
4309 1 : PageServiceCmd::PageStream(PageStreamCmd {
4310 1 : tenant_id,
4311 1 : timeline_id,
4312 1 : protocol_version: PagestreamProtocolVersion::V2,
4313 1 : })
4314 : );
4315 1 : let cmd = PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id}")).unwrap();
4316 1 : assert_eq!(
4317 : cmd,
4318 1 : PageServiceCmd::BaseBackup(BaseBackupCmd {
4319 1 : tenant_id,
4320 1 : timeline_id,
4321 1 : lsn: None,
4322 1 : gzip: false,
4323 1 : replica: false
4324 1 : })
4325 : );
4326 1 : let cmd =
4327 1 : PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id} --gzip")).unwrap();
4328 1 : assert_eq!(
4329 : cmd,
4330 1 : PageServiceCmd::BaseBackup(BaseBackupCmd {
4331 1 : tenant_id,
4332 1 : timeline_id,
4333 1 : lsn: None,
4334 1 : gzip: true,
4335 1 : replica: false
4336 1 : })
4337 : );
4338 1 : let cmd =
4339 1 : PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id} latest")).unwrap();
4340 1 : assert_eq!(
4341 : cmd,
4342 1 : PageServiceCmd::BaseBackup(BaseBackupCmd {
4343 1 : tenant_id,
4344 1 : timeline_id,
4345 1 : lsn: None,
4346 1 : gzip: false,
4347 1 : replica: false
4348 1 : })
4349 : );
4350 1 : let cmd = PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id} 0/16ABCDE"))
4351 1 : .unwrap();
4352 1 : assert_eq!(
4353 : cmd,
4354 1 : PageServiceCmd::BaseBackup(BaseBackupCmd {
4355 1 : tenant_id,
4356 1 : timeline_id,
4357 1 : lsn: Some(Lsn::from_str("0/16ABCDE").unwrap()),
4358 1 : gzip: false,
4359 1 : replica: false
4360 1 : })
4361 : );
4362 1 : let cmd = PageServiceCmd::parse(&format!(
4363 1 : "basebackup {tenant_id} {timeline_id} --replica --gzip"
4364 1 : ))
4365 1 : .unwrap();
4366 1 : assert_eq!(
4367 : cmd,
4368 1 : PageServiceCmd::BaseBackup(BaseBackupCmd {
4369 1 : tenant_id,
4370 1 : timeline_id,
4371 1 : lsn: None,
4372 1 : gzip: true,
4373 1 : replica: true
4374 1 : })
4375 : );
4376 1 : let cmd = PageServiceCmd::parse(&format!(
4377 1 : "basebackup {tenant_id} {timeline_id} 0/16ABCDE --replica --gzip"
4378 1 : ))
4379 1 : .unwrap();
4380 1 : assert_eq!(
4381 : cmd,
4382 1 : PageServiceCmd::BaseBackup(BaseBackupCmd {
4383 1 : tenant_id,
4384 1 : timeline_id,
4385 1 : lsn: Some(Lsn::from_str("0/16ABCDE").unwrap()),
4386 1 : gzip: true,
4387 1 : replica: true
4388 1 : })
4389 : );
4390 1 : let cmd = PageServiceCmd::parse(&format!("fullbackup {tenant_id} {timeline_id}")).unwrap();
4391 1 : assert_eq!(
4392 : cmd,
4393 1 : PageServiceCmd::FullBackup(FullBackupCmd {
4394 1 : tenant_id,
4395 1 : timeline_id,
4396 1 : lsn: None,
4397 1 : prev_lsn: None
4398 1 : })
4399 : );
4400 1 : let cmd = PageServiceCmd::parse(&format!(
4401 1 : "fullbackup {tenant_id} {timeline_id} 0/16ABCDE 0/16ABCDF"
4402 1 : ))
4403 1 : .unwrap();
4404 1 : assert_eq!(
4405 : cmd,
4406 1 : PageServiceCmd::FullBackup(FullBackupCmd {
4407 1 : tenant_id,
4408 1 : timeline_id,
4409 1 : lsn: Some(Lsn::from_str("0/16ABCDE").unwrap()),
4410 1 : prev_lsn: Some(Lsn::from_str("0/16ABCDF").unwrap()),
4411 1 : })
4412 : );
4413 1 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
4414 1 : let cmd = PageServiceCmd::parse(&format!(
4415 1 : "lease lsn {tenant_shard_id} {timeline_id} 0/16ABCDE"
4416 1 : ))
4417 1 : .unwrap();
4418 1 : assert_eq!(
4419 : cmd,
4420 1 : PageServiceCmd::LeaseLsn(LeaseLsnCmd {
4421 1 : tenant_shard_id,
4422 1 : timeline_id,
4423 1 : lsn: Lsn::from_str("0/16ABCDE").unwrap(),
4424 1 : })
4425 : );
4426 1 : let tenant_shard_id = TenantShardId::split(&tenant_shard_id, ShardCount(8))[1];
4427 1 : let cmd = PageServiceCmd::parse(&format!(
4428 1 : "lease lsn {tenant_shard_id} {timeline_id} 0/16ABCDE"
4429 1 : ))
4430 1 : .unwrap();
4431 1 : assert_eq!(
4432 : cmd,
4433 1 : PageServiceCmd::LeaseLsn(LeaseLsnCmd {
4434 1 : tenant_shard_id,
4435 1 : timeline_id,
4436 1 : lsn: Lsn::from_str("0/16ABCDE").unwrap(),
4437 1 : })
4438 : );
4439 1 : let cmd = PageServiceCmd::parse("set a = b").unwrap();
4440 1 : assert_eq!(cmd, PageServiceCmd::Set);
4441 1 : let cmd = PageServiceCmd::parse("SET foo").unwrap();
4442 1 : assert_eq!(cmd, PageServiceCmd::Set);
4443 1 : }
4444 :
4445 : #[test]
4446 1 : fn pageservice_cmd_err_handling() {
4447 1 : let tenant_id = TenantId::generate();
4448 1 : let timeline_id = TimelineId::generate();
4449 1 : let cmd = PageServiceCmd::parse("unknown_command");
4450 1 : assert!(cmd.is_err());
4451 1 : let cmd = PageServiceCmd::parse("pagestream_v2");
4452 1 : assert!(cmd.is_err());
4453 1 : let cmd = PageServiceCmd::parse(&format!("pagestream_v2 {tenant_id}xxx"));
4454 1 : assert!(cmd.is_err());
4455 1 : let cmd = PageServiceCmd::parse(&format!("pagestream_v2 {tenant_id}xxx {timeline_id}xxx"));
4456 1 : assert!(cmd.is_err());
4457 1 : let cmd = PageServiceCmd::parse(&format!(
4458 1 : "basebackup {tenant_id} {timeline_id} --gzip --gzip"
4459 1 : ));
4460 1 : assert!(cmd.is_err());
4461 1 : let cmd = PageServiceCmd::parse(&format!(
4462 1 : "basebackup {tenant_id} {timeline_id} --gzip --unknown"
4463 1 : ));
4464 1 : assert!(cmd.is_err());
4465 1 : let cmd = PageServiceCmd::parse(&format!(
4466 1 : "basebackup {tenant_id} {timeline_id} --gzip 0/16ABCDE"
4467 1 : ));
4468 1 : assert!(cmd.is_err());
4469 1 : let cmd = PageServiceCmd::parse(&format!("lease {tenant_id} {timeline_id} gzip 0/16ABCDE"));
4470 1 : assert!(cmd.is_err());
4471 1 : }
4472 :
4473 : #[test]
4474 1 : fn test_parse_options() {
4475 1 : let (config, has_error) = parse_options(" -c neon.compute_mode=primary ");
4476 1 : assert!(!has_error);
4477 1 : assert_eq!(
4478 : config,
4479 1 : vec![("neon.compute_mode".to_string(), "primary".to_string())]
4480 : );
4481 :
4482 1 : let (config, has_error) = parse_options(" -c neon.compute_mode=primary -c foo=bar ");
4483 1 : assert!(!has_error);
4484 1 : assert_eq!(
4485 : config,
4486 1 : vec![
4487 1 : ("neon.compute_mode".to_string(), "primary".to_string()),
4488 1 : ("foo".to_string(), "bar".to_string()),
4489 : ]
4490 : );
4491 :
4492 1 : let (config, has_error) = parse_options(" -c neon.compute_mode=primary -cfoo=bar");
4493 1 : assert!(!has_error);
4494 1 : assert_eq!(
4495 : config,
4496 1 : vec![
4497 1 : ("neon.compute_mode".to_string(), "primary".to_string()),
4498 1 : ("foo".to_string(), "bar".to_string()),
4499 : ]
4500 : );
4501 :
4502 1 : let (_, has_error) = parse_options("-c");
4503 1 : assert!(has_error);
4504 :
4505 1 : let (_, has_error) = parse_options("-c foo=bar -c -c");
4506 1 : assert!(has_error);
4507 :
4508 1 : let (_, has_error) = parse_options(" ");
4509 1 : assert!(!has_error);
4510 :
4511 1 : let (_, has_error) = parse_options(" -c neon.compute_mode");
4512 1 : assert!(has_error);
4513 1 : }
4514 : }
|
