sped up the --all option by 2x to 16x by using git cat-file --buffer

This assumes that no location log files will have a newline or carriage
return in their name. catObjectStream skips any such files due to
cat-file not supporting them.

Keys have been prevented from containing newlines since 2011,
commit 480495beb4. If some old repo
had a key with a newline in it, --all will just skip processing that key.
Other things, like .git/annex/unused files certianly assume no newlines in
keys too, and AFAICR, such keys never actually worked.

Carriage return is escaped by preSanitizeKeyName since 2013. WORM keys
generated before that point could perhaps contain a CR. (URL probably not,
http probably doesn't support an URL with a raw CR in it.) So, added
a warning in fsck about such keys. Although, fsck --all will naturally
skip them, so won't be able to warn about them. Not entirely
satisfactory, but I'll bet there are not really any such keys in
existence.

Thanks to Lukey for finding this optimisation.
This commit is contained in:
Joey Hess 2020-07-07 13:46:45 -04:00
parent 98e2e3cb9c
commit d010ab04be
No known key found for this signature in database
GPG key ID: DB12DB0FF05F8F38
5 changed files with 162 additions and 24 deletions

View file

@ -27,7 +27,7 @@ backend = Backend
, getKey = Just keyValue , getKey = Just keyValue
, verifyKeyContent = Nothing , verifyKeyContent = Nothing
, canUpgradeKey = Just needsUpgrade , canUpgradeKey = Just needsUpgrade
, fastMigrate = Just removeSpaces , fastMigrate = Just removeProblemChars
, isStableKey = const True , isStableKey = const True
} }
@ -48,12 +48,13 @@ keyValue source _ = do
, keyMtime = Just $ modificationTime stat , keyMtime = Just $ modificationTime stat
} }
{- Old WORM keys could contain spaces, and can be upgraded to remove them. -} {- Old WORM keys could contain spaces and carriage returns,
- and can be upgraded to remove them. -}
needsUpgrade :: Key -> Bool needsUpgrade :: Key -> Bool
needsUpgrade key = ' ' `S8.elem` fromKey keyName key needsUpgrade key = any (`S8.elem` fromKey keyName key) [' ', '\r']
removeSpaces :: Key -> Backend -> AssociatedFile -> Annex (Maybe Key) removeProblemChars :: Key -> Backend -> AssociatedFile -> Annex (Maybe Key)
removeSpaces oldkey newbackend _ removeProblemChars oldkey newbackend _
| migratable = return $ Just $ alterKey oldkey $ \d -> d | migratable = return $ Just $ alterKey oldkey $ \d -> d
{ keyName = encodeBS $ reSanitizeKeyName $ decodeBS $ keyName d } { keyName = encodeBS $ reSanitizeKeyName $ decodeBS $ keyName d }
| otherwise = return Nothing | otherwise = return Nothing

View file

@ -27,6 +27,12 @@ git-annex (8.20200618) UNRELEASED; urgency=medium
git-annex sync --content --all gets 20% faster. git-annex sync --content --all gets 20% faster.
* Fix a recently introduced bug that could cause a "fork: resource exhausted" * Fix a recently introduced bug that could cause a "fork: resource exhausted"
after getting several thousand files. after getting several thousand files.
* Sped up the --all option by 2x to 16x by using git cat-file --buffer.
Thanks to Lukey for finding this optimisation.
* fsck: Detect if WORM keys contain a carriage return, and recommend
upgrading to fix the problem, because the --all optimisation above
skips any such keys that might exist (git-annex could have maybe
created such keys back in 2013).
-- Joey Hess <id@joeyh.name> Thu, 18 Jun 2020 12:21:14 -0400 -- Joey Hess <id@joeyh.name> Thu, 18 Jun 2020 12:21:14 -0400

View file

@ -22,16 +22,19 @@ import qualified Git.LsTree as LsTree
import Git.FilePath import Git.FilePath
import qualified Limit import qualified Limit
import CmdLine.GitAnnex.Options import CmdLine.GitAnnex.Options
import Logs.Location import Logs
import Logs.Unused import Logs.Unused
import Types.Transfer import Types.Transfer
import Logs.Transfer import Logs.Transfer
import Remote.List import Remote.List
import qualified Remote import qualified Remote
import Annex.CatFile import Annex.CatFile
import Git.CatFile (catObjectStream)
import Annex.CurrentBranch import Annex.CurrentBranch
import Annex.Content import Annex.Content
import Annex.InodeSentinal import Annex.InodeSentinal
import qualified Annex.Branch
import qualified Annex.BranchState
import qualified Database.Keys import qualified Database.Keys
import qualified Utility.RawFilePath as R import qualified Utility.RawFilePath as R
@ -180,26 +183,55 @@ withKeyOptions' ko auto mkkeyaction fallbackaction params = do
giveup "Cannot use --auto in a bare repository" giveup "Cannot use --auto in a bare repository"
case (null params, ko) of case (null params, ko) of
(True, Nothing) (True, Nothing)
| bare -> noauto $ runkeyaction finishCheck loggedKeys | bare -> noauto runallkeys
| otherwise -> fallbackaction params | otherwise -> fallbackaction params
(False, Nothing) -> fallbackaction params (False, Nothing) -> fallbackaction params
(True, Just WantAllKeys) -> noauto $ runkeyaction finishCheck loggedKeys (True, Just WantAllKeys) -> noauto runallkeys
(True, Just WantUnusedKeys) -> noauto $ runkeyaction (pure . Just) unusedKeys' (True, Just WantUnusedKeys) -> noauto $ runkeyaction unusedKeys'
(True, Just WantFailedTransfers) -> noauto runfailedtransfers (True, Just WantFailedTransfers) -> noauto runfailedtransfers
(True, Just (WantSpecificKey k)) -> noauto $ runkeyaction (pure . Just) (return [k]) (True, Just (WantSpecificKey k)) -> noauto $ runkeyaction (return [k])
(True, Just WantIncompleteKeys) -> noauto $ runkeyaction (pure . Just) incompletekeys (True, Just WantIncompleteKeys) -> noauto $ runkeyaction incompletekeys
(True, Just (WantBranchKeys bs)) -> noauto $ runbranchkeys bs (True, Just (WantBranchKeys bs)) -> noauto $ runbranchkeys bs
(False, Just _) -> giveup "Can only specify one of file names, --all, --branch, --unused, --failed, --key, or --incomplete" (False, Just _) -> giveup "Can only specify one of file names, --all, --branch, --unused, --failed, --key, or --incomplete"
where where
noauto a noauto a
| auto = giveup "Cannot use --auto with --all or --branch or --unused or --key or --incomplete" | auto = giveup "Cannot use --auto with --all or --branch or --unused or --key or --incomplete"
| otherwise = a | otherwise = a
incompletekeys = staleKeysPrune gitAnnexTmpObjectDir True incompletekeys = staleKeysPrune gitAnnexTmpObjectDir True
runkeyaction checker getks = do
-- List all location log files on the git-annex branch,
-- and use those to get keys. Pass through cat-file
-- to get the contents of the location logs, and pre-cache
-- those. This significantly speeds up typical operations
-- that need to look at the location log for each key.
runallkeys = do
keyaction <- mkkeyaction
config <- Annex.getGitConfig
g <- Annex.gitRepo
void Annex.Branch.update
(l, cleanup) <- inRepo $ LsTree.lsTree
LsTree.LsTreeRecursive
Annex.Branch.fullname
let getk = locationLogFileKey config . getTopFilePath
let go reader = liftIO reader >>= \case
Nothing -> return ()
Just (f, content) -> do
case getk f of
Just k -> do
Annex.BranchState.setCache (getTopFilePath f) content
keyaction (k, mkActionItem k)
Nothing -> return ()
go reader
catObjectStream l (isJust . getk . LsTree.file) g go
liftIO $ void cleanup
runkeyaction getks = do
keyaction <- mkkeyaction keyaction <- mkkeyaction
ks <- getks ks <- getks
forM_ ks $ checker >=> maybe noop forM_ ks $ \k -> keyaction (k, mkActionItem k)
(\k -> keyaction (k, mkActionItem k))
runbranchkeys bs = do runbranchkeys bs = do
keyaction <- mkkeyaction keyaction <- mkkeyaction
forM_ bs $ \b -> do forM_ bs $ \b -> do
@ -211,6 +243,7 @@ withKeyOptions' ko auto mkkeyaction fallbackaction params = do
in keyaction (k, bfp) in keyaction (k, bfp)
unlessM (liftIO cleanup) $ unlessM (liftIO cleanup) $
error ("git ls-tree " ++ Git.fromRef b ++ " failed") error ("git ls-tree " ++ Git.fromRef b ++ " failed")
runfailedtransfers = do runfailedtransfers = do
keyaction <- mkkeyaction keyaction <- mkkeyaction
rs <- remoteList rs <- remoteList

View file

@ -19,6 +19,7 @@ module Git.CatFile (
catObject, catObject,
catObjectDetails, catObjectDetails,
catObjectMetaData, catObjectMetaData,
catObjectStream,
) where ) where
import System.IO import System.IO
@ -33,6 +34,7 @@ import Data.Char
import Numeric import Numeric
import System.Posix.Types import System.Posix.Types
import Text.Read import Text.Read
import Control.Concurrent.Async
import Common import Common
import Git import Git
@ -42,8 +44,10 @@ import Git.Command
import Git.Types import Git.Types
import Git.FilePath import Git.FilePath
import Git.HashObject import Git.HashObject
import qualified Git.LsTree as LsTree
import qualified Utility.CoProcess as CoProcess import qualified Utility.CoProcess as CoProcess
import Utility.Tuple import Utility.Tuple
import Control.Monad.IO.Class (MonadIO)
data CatFileHandle = CatFileHandle data CatFileHandle = CatFileHandle
{ catFileProcess :: CoProcess.CoProcessHandle { catFileProcess :: CoProcess.CoProcessHandle
@ -57,7 +61,7 @@ catFileStart = catFileStart' True
catFileStart' :: Bool -> Repo -> IO CatFileHandle catFileStart' :: Bool -> Repo -> IO CatFileHandle
catFileStart' restartable repo = CatFileHandle catFileStart' restartable repo = CatFileHandle
<$> startp "--batch" <$> startp "--batch"
<*> startp "--batch-check=%(objectname) %(objecttype) %(objectsize)" <*> startp ("--batch-check=" ++ batchFormat)
<*> pure repo <*> pure repo
where where
startp p = gitCoProcessStart restartable startp p = gitCoProcessStart restartable
@ -65,6 +69,9 @@ catFileStart' restartable repo = CatFileHandle
, Param p , Param p
] repo ] repo
batchFormat :: String
batchFormat = "%(objectname) %(objecttype) %(objectsize)"
catFileStop :: CatFileHandle -> IO () catFileStop :: CatFileHandle -> IO ()
catFileStop h = do catFileStop h = do
CoProcess.stop (catFileProcess h) CoProcess.stop (catFileProcess h)
@ -88,18 +95,12 @@ catObjectDetails :: CatFileHandle -> Ref -> IO (Maybe (L.ByteString, Sha, Object
catObjectDetails h object = query (catFileProcess h) object newlinefallback $ \from -> do catObjectDetails h object = query (catFileProcess h) object newlinefallback $ \from -> do
header <- S8.hGetLine from header <- S8.hGetLine from
case parseResp object header of case parseResp object header of
Just (ParsedResp sha objtype size) -> do Just r@(ParsedResp sha objtype _size) -> do
content <- S.hGet from (fromIntegral size) content <- readObjectContent from r
eatchar '\n' from return $ Just (content, sha, objtype)
return $ Just (L.fromChunks [content], sha, objtype)
Just DNE -> return Nothing Just DNE -> return Nothing
Nothing -> error $ "unknown response from git cat-file " ++ show (header, object) Nothing -> error $ "unknown response from git cat-file " ++ show (header, object)
where where
eatchar expected from = do
c <- hGetChar from
when (c /= expected) $
error $ "missing " ++ (show expected) ++ " from git cat-file"
-- Slow fallback path for filenames containing newlines. -- Slow fallback path for filenames containing newlines.
newlinefallback = queryObjectType object (gitRepo h) >>= \case newlinefallback = queryObjectType object (gitRepo h) >>= \case
Nothing -> return Nothing Nothing -> return Nothing
@ -113,6 +114,18 @@ catObjectDetails h object = query (catFileProcess h) object newlinefallback $ \f
(gitRepo h) (gitRepo h)
return (Just (content, sha, objtype)) return (Just (content, sha, objtype))
readObjectContent :: Handle -> ParsedResp -> IO L.ByteString
readObjectContent h (ParsedResp _ _ size) = do
content <- S.hGet h (fromIntegral size)
eatchar '\n'
return (L.fromChunks [content])
where
eatchar expected = do
c <- hGetChar h
when (c /= expected) $
error $ "missing " ++ (show expected) ++ " from git cat-file"
readObjectContent _ DNE = error "internal"
{- Gets the size and type of an object, without reading its content. -} {- Gets the size and type of an object, without reading its content. -}
catObjectMetaData :: CatFileHandle -> Ref -> IO (Maybe (Sha, FileSize, ObjectType)) catObjectMetaData :: CatFileHandle -> Ref -> IO (Maybe (Sha, FileSize, ObjectType))
catObjectMetaData h object = query (checkFileProcess h) object newlinefallback $ \from -> do catObjectMetaData h object = query (checkFileProcess h) object newlinefallback $ \from -> do
@ -266,3 +279,69 @@ parseCommit b = Commit
sp = fromIntegral (ord ' ') sp = fromIntegral (ord ' ')
lt = fromIntegral (ord '<') lt = fromIntegral (ord '<')
gt = fromIntegral (ord '>') gt = fromIntegral (ord '>')
{- Uses cat-file to stream the contents of the files listed by lstree
- as efficiently as possible. This is much faster than querying it
- repeatedly per file.
-
- Any files with a newline or carriage return in their name will be
- skipped, because the interface does not support them.
-}
catObjectStream
:: (MonadMask m, MonadIO m)
=> [LsTree.TreeItem]
-> (LsTree.TreeItem -> Bool)
-> Repo
-> (IO (Maybe (TopFilePath, L.ByteString)) -> m ())
-> m ()
catObjectStream l want repo a = assertLocal repo $
bracketIO start stop $ \(_, _, hout, _) -> a (reader hout)
where
feeder h = do
forM_ l $ \ti ->
when (want ti) $ do
let f = getTopFilePath (LsTree.file ti)
-- skip files with newlines or carriage returns
unless (any (`S8.elem` f) ['\n', '\r']) $
S8.hPutStrLn h $
fromRef' (LsTree.sha ti) <> " " <> f
hClose h
reader h = ifM (hIsEOF h)
( return Nothing
, do
resp <- S8.hGetLine h
case eitherToMaybe $ A.parseOnly respparser resp of
Nothing -> error $ "unknown response from git cat-file " ++ show resp
Just (r, f) -> do
content <- readObjectContent h r
return (Just (asTopFilePath f, content))
)
params =
[ Param "cat-file"
-- %(rest) is used to feed the filename through
-- cat-file; it will be at the end of the response
, Param ("--batch=" ++ batchFormat ++ " %(rest)")
, Param "--buffer"
]
respparser = (,)
<$> respParser
<* A8.char ' '
<*> A.takeByteString
start = do
let p = gitCreateProcess params repo
(Just hin, Just hout, _, pid) <- createProcess p
{ std_in = CreatePipe
, std_out = CreatePipe
}
f <- async (feeder hin)
return (f, hin, hout, pid)
stop (f, hin, hout, pid) = do
cancel f
hClose hin
hClose hout
void $ checkSuccessProcess pid

View file

@ -0,0 +1,19 @@
[[!comment format=mdwn
username="joey"
subject="""comment 10"""
date="2020-07-07T17:13:46Z"
content="""
Wow, I implemented your --buffer trick, and `get --all`
is over 2x faster. `sync --content --all` somewhat less, but still
another decent improvement there. (cold cache timings)
And some warm cache times are *much* faster than my cold cache benchmarks.
`get --all` is 17x faster in a 10k file repo, which makes it only 3x slower
than `get` without --all.
I think I will still leave this open because it's still worth considering
sqlite caching or finding a way to speed up the second sync --all pass...
but would be interested to know how your use case is improved now.
Please feel free to find optimisations anytime, I really appreciate it.
"""]]