git-annex/Annex/Transfer.hs
Joey Hess 579d9b60c1
improve concurrency of move/copy --from --to
Use separate stages for download and upload. In the common case where
it downloads the file from one remote and then uploads to the other,
those are by far the most expensive operations, and there's a decent
chance the two remotes bottleneck on different resources.

Suppose it's being run with -J2 and a bunch of 10 mb files. Two threads
will be started both downloading from the src remote. They will probably
finish at the same time. Then two threads will be started uploading to
the dst remote. They will probably take the same time as well. Before
this change, it would alternate back and forth, bottlenecking on src and dst.
With this change, as soon as the two threads start uploading to dst, two
more threads are able to start, downloading from src. So bandwidth to
both remotes is saturated more often.

Other commands that use transferStages only send in one direction at a
time. So the worker threads for the other direction will sit idle, and
there will be no change in their behavior.

Sponsored-by: Dartmouth College's DANDI project
2023-01-24 13:59:39 -04:00

407 lines
14 KiB
Haskell

{- git-annex transfers
-
- Copyright 2012-2021 Joey Hess <id@joeyh.name>
-
- Licensed under the GNU AGPL version 3 or higher.
-}
{-# LANGUAGE CPP, BangPatterns #-}
module Annex.Transfer (
module X,
upload,
upload',
alwaysUpload,
download,
download',
runTransfer,
alwaysRunTransfer,
noRetry,
stdRetry,
pickRemote,
) where
import Annex.Common
import qualified Annex
import Logs.Transfer as X
import Types.Transfer as X
import Annex.Notification as X
import Annex.Content
import Annex.Perms
import Annex.Action
import Utility.Metered
import Utility.ThreadScheduler
import Annex.LockPool
import Types.Key
import qualified Types.Remote as Remote
import Types.Concurrency
import Annex.Concurrent
import Types.WorkerPool
import Annex.WorkerPool
import Annex.TransferrerPool
import Annex.StallDetection
import Backend (isCryptographicallySecure)
import Types.StallDetection
import qualified Utility.RawFilePath as R
import Control.Concurrent
import Control.Concurrent.Async
import Control.Concurrent.STM hiding (retry)
import qualified Data.Map.Strict as M
import qualified System.FilePath.ByteString as P
import Data.Ord
-- Upload, supporting canceling detected stalls.
upload :: Remote -> Key -> AssociatedFile -> RetryDecider -> NotifyWitness -> Annex Bool
upload r key f d witness =
case remoteAnnexStallDetection (Remote.gitconfig r) of
Nothing -> go (Just ProbeStallDetection)
Just StallDetectionDisabled -> go Nothing
Just sd -> runTransferrer sd r key f d Upload witness
where
go sd = upload' (Remote.uuid r) key f sd d (action . Remote.storeKey r key f) witness
-- Upload, not supporting canceling detected stalls
upload' :: Observable v => UUID -> Key -> AssociatedFile -> Maybe StallDetection -> RetryDecider -> (MeterUpdate -> Annex v) -> NotifyWitness -> Annex v
upload' u key f sd d a _witness = guardHaveUUID u $
runTransfer (Transfer Upload u (fromKey id key)) f sd d a
alwaysUpload :: Observable v => UUID -> Key -> AssociatedFile -> Maybe StallDetection -> RetryDecider -> (MeterUpdate -> Annex v) -> NotifyWitness -> Annex v
alwaysUpload u key f sd d a _witness = guardHaveUUID u $
alwaysRunTransfer (Transfer Upload u (fromKey id key)) f sd d a
-- Download, supporting canceling detected stalls.
download :: Remote -> Key -> AssociatedFile -> RetryDecider -> NotifyWitness -> Annex Bool
download r key f d witness =
case remoteAnnexStallDetection (Remote.gitconfig r) of
Nothing -> go (Just ProbeStallDetection)
Just StallDetectionDisabled -> go Nothing
Just sd -> runTransferrer sd r key f d Download witness
where
go sd = getViaTmp (Remote.retrievalSecurityPolicy r) vc key f $ \dest ->
download' (Remote.uuid r) key f sd d (go' dest) witness
go' dest p = verifiedAction $
Remote.retrieveKeyFile r key f (fromRawFilePath dest) p vc
vc = Remote.RemoteVerify r
-- Download, not supporting canceling detected stalls.
download' :: Observable v => UUID -> Key -> AssociatedFile -> Maybe StallDetection -> RetryDecider -> (MeterUpdate -> Annex v) -> NotifyWitness -> Annex v
download' u key f sd d a _witness = guardHaveUUID u $
runTransfer (Transfer Download u (fromKey id key)) f sd d a
guardHaveUUID :: Observable v => UUID -> Annex v -> Annex v
guardHaveUUID u a
| u == NoUUID = return observeFailure
| otherwise = a
{- Runs a transfer action. Creates and locks the lock file while the
- action is running, and stores info in the transfer information
- file.
-
- If the transfer action returns False, the transfer info is
- left in the failedTransferDir.
-
- If the transfer is already in progress, returns False.
-
- An upload can be run from a read-only filesystem, and in this case
- no transfer information or lock file is used.
-
- Cannot cancel stalls, but when a likely stall is detected,
- suggests to the user that they enable stall detection handling.
-}
runTransfer :: Observable v => Transfer -> AssociatedFile -> Maybe StallDetection -> RetryDecider -> (MeterUpdate -> Annex v) -> Annex v
runTransfer = runTransfer' False
{- Like runTransfer, but ignores any existing transfer lock file for the
- transfer, allowing re-running a transfer that is already in progress.
-}
alwaysRunTransfer :: Observable v => Transfer -> AssociatedFile -> Maybe StallDetection -> RetryDecider -> (MeterUpdate -> Annex v) -> Annex v
alwaysRunTransfer = runTransfer' True
runTransfer' :: Observable v => Bool -> Transfer -> AssociatedFile -> Maybe StallDetection -> RetryDecider -> (MeterUpdate -> Annex v) -> Annex v
runTransfer' ignorelock t afile stalldetection retrydecider transferaction =
enteringStage (TransferStage (transferDirection t)) $
debugLocks $
preCheckSecureHashes (transferKey t) go
where
go = do
info <- liftIO $ startTransferInfo afile
(meter, tfile, createtfile, metervar) <- mkProgressUpdater t info
mode <- annexFileMode
(lck, inprogress) <- prep tfile createtfile mode
if inprogress && not ignorelock
then do
warning "transfer already in progress, or unable to take transfer lock"
return observeFailure
else do
v <- retry 0 info metervar $
detectStallsAndSuggestConfig stalldetection metervar $
transferaction meter
liftIO $ cleanup tfile lck
if observeBool v
then removeFailedTransfer t
else recordFailedTransfer t info
return v
prep :: RawFilePath -> Annex () -> FileMode -> Annex (Maybe LockHandle, Bool)
#ifndef mingw32_HOST_OS
prep tfile createtfile mode = catchPermissionDenied (const prepfailed) $ do
let lck = transferLockFile tfile
createAnnexDirectory $ P.takeDirectory lck
tryLockExclusive (Just mode) lck >>= \case
Nothing -> return (Nothing, True)
-- Since the lock file is removed in cleanup,
-- there's a race where different processes
-- may have a deleted and a new version of the same
-- lock file open. checkSaneLock guards against
-- that.
Just lockhandle -> ifM (checkSaneLock lck lockhandle)
( do
createtfile
return (Just lockhandle, False)
, do
liftIO $ dropLock lockhandle
return (Nothing, True)
)
#else
prep tfile createtfile _mode = catchPermissionDenied (const prepfailed) $ do
let lck = transferLockFile tfile
createAnnexDirectory $ P.takeDirectory lck
catchMaybeIO (liftIO $ lockExclusive lck) >>= \case
Nothing -> return (Nothing, False)
Just Nothing -> return (Nothing, True)
Just (Just lockhandle) -> do
createtfile
return (Just lockhandle, False)
#endif
prepfailed = return (Nothing, False)
cleanup _ Nothing = noop
cleanup tfile (Just lockhandle) = do
let lck = transferLockFile tfile
void $ tryIO $ R.removeLink tfile
#ifndef mingw32_HOST_OS
void $ tryIO $ R.removeLink lck
dropLock lockhandle
#else
{- Windows cannot delete the lockfile until the lock
- is closed. So it's possible to race with another
- process that takes the lock before it's removed,
- so ignore failure to remove.
-}
dropLock lockhandle
void $ tryIO $ R.removeLink lck
#endif
retry numretries oldinfo metervar run =
tryNonAsync run >>= \case
Right v
| observeBool v -> return v
| otherwise -> checkretry
Left e -> do
warning (show e)
checkretry
where
checkretry = do
b <- getbytescomplete metervar
let newinfo = oldinfo { bytesComplete = Just b }
let !numretries' = succ numretries
ifM (retrydecider numretries' oldinfo newinfo)
( retry numretries' newinfo metervar run
, return observeFailure
)
getbytescomplete metervar = liftIO $
maybe 0 fromBytesProcessed <$> readTVarIO metervar
detectStallsAndSuggestConfig :: Maybe StallDetection -> TVar (Maybe BytesProcessed) -> Annex a -> Annex a
detectStallsAndSuggestConfig Nothing _ a = a
detectStallsAndSuggestConfig sd@(Just _) metervar a =
bracket setup cleanup (const a)
where
setup = do
v <- liftIO newEmptyTMVarIO
sdt <- liftIO $ async $ detectStalls sd metervar $
void $ atomically $ tryPutTMVar v True
wt <- liftIO . async =<< forkState (warnonstall v)
return (v, sdt, wt)
cleanup (v, sdt, wt) = do
liftIO $ uninterruptibleCancel sdt
void $ liftIO $ atomically $ tryPutTMVar v False
join (liftIO (wait wt))
warnonstall v = whenM (liftIO (atomically (takeTMVar v))) $
warning "Transfer seems to have stalled. To restart stalled transfers, configure annex.stalldetection"
{- Runs a transfer using a separate process, which lets detected stalls be
- canceled. -}
runTransferrer
:: StallDetection
-> Remote
-> Key
-> AssociatedFile
-> RetryDecider
-> Direction
-> NotifyWitness
-> Annex Bool
runTransferrer sd r k afile retrydecider direction _witness =
enteringStage (TransferStage direction) $ preCheckSecureHashes k $ do
info <- liftIO $ startTransferInfo afile
go 0 info
where
go numretries info =
withTransferrer (performTransfer (Just sd) AnnexLevel id (Just r) t info) >>= \case
Right () -> return True
Left newinfo -> do
let !numretries' = succ numretries
ifM (retrydecider numretries' info newinfo)
( go numretries' newinfo
, return False
)
t = Transfer direction (Remote.uuid r) (fromKey id k)
{- Avoid download and upload of keys with insecure content when
- annex.securehashesonly is configured.
-
- This is not a security check. Even if this let the content be
- downloaded, the actual security checks would prevent the content from
- being added to the repository. The only reason this is done here is to
- avoid transferring content that's going to be rejected anyway.
-
- We assume that, if annex.securehashesonly is set and the local repo
- still contains content using an insecure hash, remotes will likewise
- tend to be configured to reject it, so Upload is also prevented.
-}
preCheckSecureHashes :: Observable v => Key -> Annex v -> Annex v
preCheckSecureHashes k a = ifM (isCryptographicallySecure k)
( a
, ifM (annexSecureHashesOnly <$> Annex.getGitConfig)
( do
warning $ "annex.securehashesonly blocked transfer of " ++ decodeBS (formatKeyVariety variety) ++ " key"
return observeFailure
, a
)
)
where
variety = fromKey keyVariety k
type NumRetries = Integer
type RetryDecider = NumRetries -> TransferInfo -> TransferInfo -> Annex Bool
{- Both retry deciders are checked together, so if one chooses to delay,
- it will always take effect. -}
combineRetryDeciders :: RetryDecider -> RetryDecider -> RetryDecider
combineRetryDeciders a b = \n old new -> do
ar <- a n old new
br <- b n old new
return (ar || br)
noRetry :: RetryDecider
noRetry _ _ _ = pure False
stdRetry :: RetryDecider
stdRetry = combineRetryDeciders forwardRetry configuredRetry
{- Keep retrying failed transfers, as long as forward progress is being
- made.
-
- Up to a point -- while some remotes can resume where the previous
- transfer left off, and so it would make sense to keep retrying forever,
- other remotes restart each transfer from the beginning, and so even if
- forward progress is being made, it's not real progress. So, retry a
- maximum of 5 times by default.
-}
forwardRetry :: RetryDecider
forwardRetry numretries old new
| fromMaybe 0 (bytesComplete old) < fromMaybe 0 (bytesComplete new) =
(numretries <=) <$> maybe globalretrycfg pure remoteretrycfg
| otherwise = return False
where
globalretrycfg = fromMaybe 5 . annexForwardRetry
<$> Annex.getGitConfig
remoteretrycfg = remoteAnnexRetry =<<
(Remote.gitconfig <$> transferRemote new)
{- Retries a number of times with growing delays in between when enabled
- by git configuration. -}
configuredRetry :: RetryDecider
configuredRetry numretries _old new = do
(maxretries, Seconds initretrydelay) <- getcfg $
Remote.gitconfig <$> transferRemote new
if numretries < maxretries
then do
let retrydelay = Seconds (initretrydelay * 2^(numretries-1))
showSideAction $ "Delaying " ++ show (fromSeconds retrydelay) ++ "s before retrying."
liftIO $ threadDelaySeconds retrydelay
return True
else return False
where
globalretrycfg = fromMaybe 0 . annexRetry
<$> Annex.getGitConfig
globalretrydelaycfg = fromMaybe (Seconds 1) . annexRetryDelay
<$> Annex.getGitConfig
getcfg Nothing = (,) <$> globalretrycfg <*> globalretrydelaycfg
getcfg (Just gc) = (,)
<$> maybe globalretrycfg return (remoteAnnexRetry gc)
<*> maybe globalretrydelaycfg return (remoteAnnexRetryDelay gc)
{- Picks a remote from the list and tries a transfer to it. If the transfer
- does not succeed, goes on to try other remotes from the list.
-
- The list should already be ordered by remote cost, and is normally
- tried in order. However, when concurrent jobs are running, they will
- be assigned different remotes of the same cost when possible. This can
- increase total transfer speed.
-}
pickRemote :: Observable v => [Remote] -> (Remote -> Annex v) -> Annex v
pickRemote l a = debugLocks $ go l =<< getConcurrency
where
go [] _ = return observeFailure
go (r:[]) _ = a r
go rs NonConcurrent = gononconcurrent rs
go rs (Concurrent n)
| n <= 1 = gononconcurrent rs
| otherwise = goconcurrent rs
go rs ConcurrentPerCpu = goconcurrent rs
gononconcurrent [] = return observeFailure
gononconcurrent (r:rs) = do
ok <- a r
if observeBool ok
then return ok
else gononconcurrent rs
goconcurrent rs = do
mv <- Annex.getRead Annex.activeremotes
active <- liftIO $ takeMVar mv
let rs' = sortBy (lessActiveFirst active) rs
goconcurrent' mv active rs'
goconcurrent' mv active [] = do
liftIO $ putMVar mv active
return observeFailure
goconcurrent' mv active (r:rs) = do
let !active' = M.insertWith (+) r 1 active
liftIO $ putMVar mv active'
let getnewactive = do
active'' <- liftIO $ takeMVar mv
let !active''' = M.update (\n -> if n > 1 then Just (n-1) else Nothing) r active''
return active'''
let removeactive = liftIO . putMVar mv =<< getnewactive
ok <- a r `onException` removeactive
if observeBool ok
then do
removeactive
return ok
else do
active'' <- getnewactive
-- Re-sort the remaining rs
-- because other threads could have
-- been assigned them in the meantime.
let rs' = sortBy (lessActiveFirst active'') rs
goconcurrent' mv active'' rs'
lessActiveFirst :: M.Map Remote Integer -> Remote -> Remote -> Ordering
lessActiveFirst active a b
| Remote.cost a == Remote.cost b = comparing (`M.lookup` active) a b
| otherwise = comparing Remote.cost a b