/***********************************************************************************************************************************
Remote command
***********************************************************************************************************************************/
void
cmdRemote(int handleRead, int handleWrite)
{
FUNCTION_LOG_VOID(logLevelDebug);
MEM_CONTEXT_TEMP_BEGIN()
{
String *name = strNewFmt(PROTOCOL_SERVICE_REMOTE "-%u", cfgOptionUInt(cfgOptProcess));
IoRead *read = ioHandleReadNew(name, handleRead, (TimeMSec)(cfgOptionDbl(cfgOptProtocolTimeout) * 1000));
ioReadOpen(read);
IoWrite *write = ioHandleWriteNew(name, handleWrite);
ioWriteOpen(write);
ProtocolServer *server = protocolServerNew(name, PROTOCOL_SERVICE_REMOTE_STR, read, write);
protocolServerHandlerAdd(server, storageRemoteProtocol);
protocolServerHandlerAdd(server, configProtocol);
// Acquire a lock if this command needs one. We'll use the noop that is always sent from the client right after the
// handshake to return an error.
volatile bool success = false;
TRY_BEGIN()
{
// Read the command. No need to parse it since we know this is the first noop.
ioReadLine(read);
// Only try the lock if this is process 0, i.e. the remote started from the main process
if (cfgOptionUInt(cfgOptProcess) == 0)
{
ConfigCommand commandId = cfgCommandId(strPtr(cfgOptionStr(cfgOptCommand)));
// Acquire a lock if this command requires a lock
if (cfgLockRemoteRequired(commandId))
lockAcquire(cfgOptionStr(cfgOptLockPath), cfgOptionStr(cfgOptStanza), cfgLockRemoteType(commandId), 0, true);
}
protocolServerResponse(server, NULL);
success = true;
}
CATCH_ANY()
{
protocolServerError(server, errorCode(), STR(errorMessage()), STR(errorStackTrace()));
}
TRY_END();
// If not successful we'll just exit
if (success)
protocolServerProcess(server);
}
MEM_CONTEXT_TEMP_END();
FUNCTION_LOG_RETURN_VOID();
}
/***********************************************************************************************************************************
Get an archive file from the repository (WAL segment, history file, etc.)
***********************************************************************************************************************************/
int
cmdArchiveGet(void)
{
FUNCTION_LOG_VOID(logLevelDebug);
// Set the result assuming the archive file will not be found
int result = 1;
MEM_CONTEXT_TEMP_BEGIN()
{
// Check the parameters
const StringList *commandParam = cfgCommandParam();
if (strLstSize(commandParam) != 2)
{
if (strLstSize(commandParam) == 0)
THROW(ParamRequiredError, "WAL segment to get required");
if (strLstSize(commandParam) == 1)
THROW(ParamRequiredError, "path to copy WAL segment required");
THROW(ParamInvalidError, "extra parameters found");
}
// Get the segment name
String *walSegment = strBase(strLstGet(commandParam, 0));
// Destination is wherever we were told to move the WAL segment
const String *walDestination =
walPath(strLstGet(commandParam, 1), cfgOptionStr(cfgOptPgPath), STR(cfgCommandName(cfgCommand())));
// Async get can only be performed on WAL segments, history or other files must use synchronous mode
if (cfgOptionBool(cfgOptArchiveAsync) && walIsSegment(walSegment))
{
bool found = false; // Has the WAL segment been found yet?
bool queueFull = false; // Is the queue half or more full?
bool forked = false; // Has the async process been forked yet?
bool confessOnError = false; // Should we confess errors?
// Loop and wait for the WAL segment to be pushed
Wait *wait = waitNew((TimeMSec)(cfgOptionDbl(cfgOptArchiveTimeout) * MSEC_PER_SEC));
do
{
// Check for errors or missing files. For archive-get ok indicates that the process succeeded but there is no WAL
// file to download.
if (archiveAsyncStatus(archiveModeGet, walSegment, confessOnError))
{
storageRemoveP(
storageSpoolWrite(),
strNewFmt(STORAGE_SPOOL_ARCHIVE_IN "/%s" STATUS_EXT_OK, strPtr(walSegment)), .errorOnMissing = true);
break;
}
// Check if the WAL segment is already in the queue
found = storageExistsNP(storageSpool(), strNewFmt(STORAGE_SPOOL_ARCHIVE_IN "/%s", strPtr(walSegment)));
// If found then move the WAL segment to the destination directory
if (found)
{
// Source is the WAL segment in the spool queue
StorageFileRead *source = storageNewReadNP(
storageSpool(), strNewFmt(STORAGE_SPOOL_ARCHIVE_IN "/%s", strPtr(walSegment)));
// A move will be attempted but if the spool queue and the WAL path are on different file systems then a copy
// will be performed instead.
//
// It looks scary that we are disabling syncs and atomicity (in case we need to copy intead of move) but this
// is safe because if the system crashes Postgres will not try to reuse a restored WAL segment but will instead
// request it again using the restore_command. In the case of a move this hardly matters since path syncs are
// cheap but if a copy is required we could save a lot of writes.
StorageFileWrite *destination = storageNewWriteP(
storageLocalWrite(), walDestination, .noCreatePath = true, .noSyncFile = true, .noSyncPath = true,
.noAtomic = true);
// Move (or copy if required) the file
storageMoveNP(storageSpoolWrite(), source, destination);
// Return success
result = 0;
// Get a list of WAL segments left in the queue
StringList *queue = storageListP(
storageSpool(), STORAGE_SPOOL_ARCHIVE_IN_STR, .expression = WAL_SEGMENT_REGEXP_STR);
if (strLstSize(queue) > 0)
{
// Get size of the WAL segment
uint64_t walSegmentSize = storageInfoNP(storageLocal(), walDestination).size;
// Use WAL segment size to estimate queue size and determine if the async process should be launched
queueFull = strLstSize(queue) * walSegmentSize > cfgOptionUInt64(cfgOptArchiveGetQueueMax) / 2;
}
}
// If the WAL segment has not already been found then start the async process to get it. There's no point in
// forking the async process off more than once so track that as well. Use an archive lock to prevent forking if
// the async process was launched by another process.
if (!forked && (!found || !queueFull) &&
lockAcquire(cfgOptionStr(cfgOptLockPath), cfgOptionStr(cfgOptStanza), cfgLockType(), 0, false))
{
// Get control info
PgControl pgControl = pgControlFromFile(cfgOptionStr(cfgOptPgPath));
// Create the queue
storagePathCreateNP(storageSpoolWrite(), STORAGE_SPOOL_ARCHIVE_IN_STR);
// The async process should not output on the console at all
KeyValue *optionReplace = kvNew();
kvPut(optionReplace, VARSTR(CFGOPT_LOG_LEVEL_CONSOLE_STR), VARSTRDEF("off"));
kvPut(optionReplace, VARSTR(CFGOPT_LOG_LEVEL_STDERR_STR), VARSTRDEF("off"));
// Generate command options
StringList *commandExec = cfgExecParam(cfgCmdArchiveGetAsync, optionReplace);
strLstInsert(commandExec, 0, cfgExe());
// Clean the current queue using the list of WAL that we ideally want in the queue. queueNeed()
// will return the list of WAL needed to fill the queue and this will be passed to the async process.
const StringList *queue = queueNeed(
walSegment, found, cfgOptionUInt64(cfgOptArchiveGetQueueMax), pgControl.walSegmentSize,
pgControl.version);
for (unsigned int queueIdx = 0; queueIdx < strLstSize(queue); queueIdx++)
strLstAdd(commandExec, strLstGet(queue, queueIdx));
// Release the lock so the child process can acquire it
lockRelease(true);
// Fork off the async process
if (forkSafe() == 0)
{
// Disable logging and close log file
logClose();
// Detach from parent process
forkDetach();
// Execute the binary. This statement will not return if it is successful.
THROW_ON_SYS_ERROR(
execvp(strPtr(cfgExe()), (char ** const)strLstPtr(commandExec)) == -1,
ExecuteError, "unable to execute '" CFGCMD_ARCHIVE_GET_ASYNC "'");
}
// Mark the async process as forked so it doesn't get forked again. A single run of the async process should be
// enough to do the job, running it again won't help anything.
forked = true;
}
// Exit loop if WAL was found
if (found)
break;
// Now that the async process has been launched, confess any errors that are found
confessOnError = true;
}
while (waitMore(wait));
}
// Else perform synchronous get
else
{
