Line data Source code
1 : use anyhow::{bail, Result};
2 : use camino::Utf8PathBuf;
3 : use postgres_ffi::{MAX_SEND_SIZE, WAL_SEGMENT_SIZE};
4 : use std::sync::Arc;
5 : use tokio::{
6 : fs::OpenOptions,
7 : io::{AsyncSeekExt, AsyncWriteExt},
8 : };
9 : use tracing::{info, warn};
10 : use utils::{id::TenantTimelineId, lsn::Lsn};
11 :
12 : use crate::{
13 : control_file::FileStorage,
14 : state::TimelinePersistentState,
15 : timeline::{TimelineError, WalResidentTimeline},
16 : timelines_global_map::{create_temp_timeline_dir, validate_temp_timeline},
17 : wal_backup::copy_s3_segments,
18 : wal_storage::{wal_file_paths, WalReader},
19 : GlobalTimelines,
20 : };
21 :
22 : // we don't want to have more than 10 segments on disk after copy, because they take space
23 : const MAX_BACKUP_LAG: u64 = 10 * WAL_SEGMENT_SIZE as u64;
24 :
25 : pub struct Request {
26 : pub source_ttid: TenantTimelineId,
27 : pub until_lsn: Lsn,
28 : pub destination_ttid: TenantTimelineId,
29 : }
30 :
31 0 : pub async fn handle_request(
32 0 : request: Request,
33 0 : global_timelines: Arc<GlobalTimelines>,
34 0 : ) -> Result<()> {
35 0 : // TODO: request.until_lsn MUST be a valid LSN, and we cannot check it :(
36 0 : // if LSN will point to the middle of a WAL record, timeline will be in "broken" state
37 0 :
38 0 : match global_timelines.get(request.destination_ttid) {
39 : // timeline already exists. would be good to check that this timeline is the copy
40 : // of the source timeline, but it isn't obvious how to do that
41 0 : Ok(_) => return Ok(()),
42 : // timeline not found, we are going to create it
43 0 : Err(TimelineError::NotFound(_)) => {}
44 : // error, probably timeline was deleted
45 0 : res => {
46 0 : res?;
47 : }
48 : }
49 :
50 0 : let source = global_timelines.get(request.source_ttid)?;
51 0 : let source_tli = source.wal_residence_guard().await?;
52 :
53 0 : let conf = &global_timelines.get_global_config();
54 0 : let ttid = request.destination_ttid;
55 :
56 0 : let (_tmp_dir, tli_dir_path) = create_temp_timeline_dir(conf, ttid).await?;
57 :
58 0 : let (mem_state, state) = source_tli.get_state().await;
59 0 : let start_lsn = state.timeline_start_lsn;
60 0 : if start_lsn == Lsn::INVALID {
61 0 : bail!("timeline is not initialized");
62 0 : }
63 0 : let backup_lsn = mem_state.backup_lsn;
64 0 :
65 0 : {
66 0 : let commit_lsn = mem_state.commit_lsn;
67 0 : let flush_lsn = source_tli.get_flush_lsn().await;
68 :
69 0 : info!(
70 0 : "collected info about source timeline: start_lsn={}, backup_lsn={}, commit_lsn={}, flush_lsn={}",
71 : start_lsn, backup_lsn, commit_lsn, flush_lsn
72 : );
73 :
74 0 : assert!(backup_lsn >= start_lsn);
75 0 : assert!(commit_lsn >= start_lsn);
76 0 : assert!(flush_lsn >= start_lsn);
77 :
78 0 : if request.until_lsn > flush_lsn {
79 0 : bail!(format!(
80 0 : "requested LSN {} is beyond the end of the timeline {}",
81 0 : request.until_lsn, flush_lsn
82 0 : ));
83 0 : }
84 0 : if request.until_lsn < start_lsn {
85 0 : bail!(format!(
86 0 : "requested LSN {} is before the start of the timeline {}",
87 0 : request.until_lsn, start_lsn
88 0 : ));
89 0 : }
90 0 :
91 0 : if request.until_lsn > commit_lsn {
92 0 : warn!("copy_timeline WAL is not fully committed");
93 0 : }
94 :
95 0 : if backup_lsn < request.until_lsn && request.until_lsn.0 - backup_lsn.0 > MAX_BACKUP_LAG {
96 : // we have a lot of segments that are not backed up. we can try to wait here until
97 : // segments will be backed up to remote storage, but it's not clear how long to wait
98 0 : bail!("too many segments are not backed up");
99 0 : }
100 0 : }
101 0 :
102 0 : let wal_seg_size = state.server.wal_seg_size as usize;
103 0 : if wal_seg_size == 0 {
104 0 : bail!("wal_seg_size is not set");
105 0 : }
106 0 :
107 0 : let first_segment = start_lsn.segment_number(wal_seg_size);
108 0 : let last_segment = request.until_lsn.segment_number(wal_seg_size);
109 :
110 0 : let new_backup_lsn = {
111 : // we can't have new backup_lsn greater than existing backup_lsn or start of the last segment
112 0 : let max_backup_lsn = backup_lsn.min(Lsn(last_segment * wal_seg_size as u64));
113 0 :
114 0 : if max_backup_lsn <= start_lsn {
115 : // probably we are starting from the first segment, which was not backed up yet.
116 : // note that start_lsn can be in the middle of the segment
117 0 : start_lsn
118 : } else {
119 : // we have some segments backed up, so we will assume all WAL below max_backup_lsn is backed up
120 0 : assert!(max_backup_lsn.segment_offset(wal_seg_size) == 0);
121 0 : max_backup_lsn
122 : }
123 : };
124 :
125 : // all previous segments will be copied inside S3
126 0 : let first_ondisk_segment = new_backup_lsn.segment_number(wal_seg_size);
127 0 : assert!(first_ondisk_segment <= last_segment);
128 0 : assert!(first_ondisk_segment >= first_segment);
129 :
130 0 : copy_s3_segments(
131 0 : wal_seg_size,
132 0 : &request.source_ttid,
133 0 : &request.destination_ttid,
134 0 : first_segment,
135 0 : first_ondisk_segment,
136 0 : )
137 0 : .await?;
138 :
139 0 : copy_disk_segments(
140 0 : &source_tli,
141 0 : wal_seg_size,
142 0 : new_backup_lsn,
143 0 : request.until_lsn,
144 0 : &tli_dir_path,
145 0 : )
146 0 : .await?;
147 :
148 0 : let mut new_state = TimelinePersistentState::new(
149 0 : &request.destination_ttid,
150 0 : state.server.clone(),
151 0 : vec![],
152 0 : request.until_lsn,
153 0 : start_lsn,
154 0 : )?;
155 0 : new_state.timeline_start_lsn = start_lsn;
156 0 : new_state.peer_horizon_lsn = request.until_lsn;
157 0 : new_state.backup_lsn = new_backup_lsn;
158 0 :
159 0 : FileStorage::create_new(&tli_dir_path, new_state.clone(), conf.no_sync).await?;
160 :
161 : // now we have a ready timeline in a temp directory
162 0 : validate_temp_timeline(conf, request.destination_ttid, &tli_dir_path).await?;
163 0 : global_timelines
164 0 : .load_temp_timeline(request.destination_ttid, &tli_dir_path, true)
165 0 : .await?;
166 :
167 0 : Ok(())
168 0 : }
169 :
170 0 : async fn copy_disk_segments(
171 0 : tli: &WalResidentTimeline,
172 0 : wal_seg_size: usize,
173 0 : start_lsn: Lsn,
174 0 : end_lsn: Lsn,
175 0 : tli_dir_path: &Utf8PathBuf,
176 0 : ) -> Result<()> {
177 0 : let mut wal_reader = tli.get_walreader(start_lsn).await?;
178 :
179 0 : let mut buf = vec![0u8; MAX_SEND_SIZE];
180 0 :
181 0 : let first_segment = start_lsn.segment_number(wal_seg_size);
182 0 : let last_segment = end_lsn.segment_number(wal_seg_size);
183 :
184 0 : for segment in first_segment..=last_segment {
185 0 : let segment_start = segment * wal_seg_size as u64;
186 0 : let segment_end = segment_start + wal_seg_size as u64;
187 0 :
188 0 : let copy_start = segment_start.max(start_lsn.0);
189 0 : let copy_end = segment_end.min(end_lsn.0);
190 0 :
191 0 : let copy_start = copy_start - segment_start;
192 0 : let copy_end = copy_end - segment_start;
193 :
194 0 : let wal_file_path = {
195 0 : let (normal, partial) = wal_file_paths(tli_dir_path, segment, wal_seg_size);
196 0 :
197 0 : if segment == last_segment {
198 0 : partial
199 : } else {
200 0 : normal
201 : }
202 : };
203 :
204 0 : write_segment(
205 0 : &mut buf,
206 0 : &wal_file_path,
207 0 : wal_seg_size as u64,
208 0 : copy_start,
209 0 : copy_end,
210 0 : &mut wal_reader,
211 0 : )
212 0 : .await?;
213 : }
214 :
215 0 : Ok(())
216 0 : }
217 :
218 0 : async fn write_segment(
219 0 : buf: &mut [u8],
220 0 : file_path: &Utf8PathBuf,
221 0 : wal_seg_size: u64,
222 0 : from: u64,
223 0 : to: u64,
224 0 : reader: &mut WalReader,
225 0 : ) -> Result<()> {
226 0 : assert!(from <= to);
227 0 : assert!(to <= wal_seg_size);
228 :
229 : #[allow(clippy::suspicious_open_options)]
230 0 : let mut file = OpenOptions::new()
231 0 : .create(true)
232 0 : .write(true)
233 0 : .open(&file_path)
234 0 : .await?;
235 :
236 : // maybe fill with zeros, as in wal_storage.rs?
237 0 : file.set_len(wal_seg_size).await?;
238 0 : file.seek(std::io::SeekFrom::Start(from)).await?;
239 :
240 0 : let mut bytes_left = to - from;
241 0 : while bytes_left > 0 {
242 0 : let len = bytes_left as usize;
243 0 : let len = len.min(buf.len());
244 0 : let len = reader.read(&mut buf[..len]).await?;
245 0 : file.write_all(&buf[..len]).await?;
246 0 : bytes_left -= len as u64;
247 : }
248 :
249 0 : file.flush().await?;
250 0 : file.sync_all().await?;
251 0 : Ok(())
252 0 : }
|
