This avoids walking the entire heap multiple times for complex
operations like readdir. This does not entirely eliminate the
observed performance regression but does dramatically reduce s3fs CPU
usage. References #935. Fixes#936.
1) Changed codes about memory leak
For memory leak, below codes are changed.
* calls malloc_trim function
* calls initializing NSS function, and adds configure
option "--enable-nss-init".
If libcurl is with NSS, s3fs initializes NSS manually.
This initializing NSS is enabled by "--enable-nss-init"
option at configure. if this option is specified, you
need "nss-devel" package.
* calls initializing libxml2(xmlInitParser).
* BIO functions have memory leak, calls CRYPTO_free_ex_data.
* changes cache structure.
* changes cache out logic to LRU.
* sets alignment for allcated memory in body data structure.
* adds ssl session into share handle. and adds nosscache option.
* deletes unused allocated memory.(bug)
* changes defaule parallel count of head request in readdir
(500->20)
* fixes some bugs.
git-svn-id: http://s3fs.googlecode.com/svn/trunk@482 df820570-a93a-0410-bd06-b72b767a4274
1) Adds new S3fsCurl class
Added new S3fsCurl class instead of directly calling curl function.
This class is lapping curl function for s3fs(AWS S3 API).
2) Changes codes about adding S3fsCurl class
Changed and deleted classes and structures which are related to curl in curl.cpp/curl.h.
Changed codes which are calling S3 API with curl in s3fs.cpp.
3) Deletes YKIES macro
Deleted YIKES macro, because this macro is used no more.
4) Changes a code
s3fs does not get good performance because s3fs copies each byte while downloading.
So that the codes is changed instead of memcpy, then s3fs performance not a little improves.
5) Fixes a bug
When s3fs renames a file, s3fs does not use the value which is specified by servicepath option.
Fixed this bug.
git-svn-id: http://s3fs.googlecode.com/svn/trunk@451 df820570-a93a-0410-bd06-b72b767a4274
No changes for logic, only changes layout of functions and valiables between a file to a file.
Adds s3fs_util.cpp/s3fs_util.h/common.h
git-svn-id: http://s3fs.googlecode.com/svn/trunk@396 df820570-a93a-0410-bd06-b72b767a4274
* This new version was made for fixing big issue about directory object.
Please be careful and review new s3fs.
==========================
List of Changes
==========================
1) Fixed bugs
Fixed some memory leak and un-freed curl handle.
Fixed codes with a bug which is not found yet.
Fixed a bug that the s3fs could not update object's mtime when the s3fs had a opened file descriptor.
Please let us know a bug, when you find new bug of a memory leak.
2) Changed codes
Changed codes of s3fs_readdir() and list_bucket() etc.
Changed codes so that the get_realpath() function returned std::string.
Changed codes about exit() function. Because the exit() function is called from many fuse callback function directly, these function called fuse_exit() function and retuned with error.
Changed codes so that the case of the characters for the "x-amz-meta" response header is ignored.
3) Added a option
Added the norenameapi option for the storage compatible with S3 without copy API.
This option is subset of nocopyapi option.
Please read man page or call with --help option.
4) Object for directory
This is very big and important change.
The object of directory is changed "dir/" instead of "dir" for being compatible with other S3 client applications.
And this version understands the object of directory which is made by old version.
If the new s3fs changes the attributes or owner/group or mtime of the directory object, the s3fs automatically changes the object from old object name("dir") to new("dir/").
If you need to change old object name("dir") to new("dir/") manually, you can use shell script(mergedir.sh) in test directory.
* About the directory object name
AWS S3 allows the object name as both "dir" and "dir/".
The s3fs before this version understood only "dir" as directory object name, but old version did not understand the "dir/" object name.
The new version understands both of "dir" and "dir/" object name.
The s3fs user needs to be care for the special situation that I mentioned later.
The new version deletes old "dir" object and makes new "dir/" object, when the user operates the directory object for changing the permission or owner/group or mtime.
This operation does on background and automatically.
If you need to merge manually, you can use shell script which is mergedir.sh in test directory.
This script runs chmod/chown/touch commands after finding a directory.
Other S3 client application makes a directory object("dir/") without meta information which is needed to understand by the s3fs, this script can add meta information for a directory object.
If this script function is insufficient for you, you can read and modify the codes by yourself.
Please use the shell script carefully because of changing the object.
If you find a bug in this script, please let me know.
* Details
** The directory object made by old version
The directory object made by old version is not understood by other S3 client application.
New s3fs version was updated for keeping compatibility with other clients.
You can use the mergedir.sh in test directory for merging from old directory object("dir") to new("dir/").
The directory object name is changed from "dir" to "dir/" after the mergedir.sh is run, this changed "dir/" object is understood by other S3 clients.
This script runs chmod/chown/chgrp/touch/etc commands against the old directory object("dir"), then new s3fs merges that directory automatically.
If you need to change directory object from old to new manually, you can do it by running these commands which change the directory attributes(mode/owner/group/mtime).
** The directory object made by new version
The directory object name made by new version is "dir/".
Because the name includes "/", other S3 client applications understand it as the directory.
I tested new directory by s3cmd/tntDrive/DragonDisk/Gladinet as other S3 clients, the result was good compatibility.
You need to know that the compatibility has small problem by the difference in specifications between clients.
And you need to be careful about that the old s3fs can not understand the directory object which made by new s3fs.
You should change all s3fs which accesses same bucket.
** The directory object made by other S3 client application
Because the object is determined as a directory by the s3fs, the s3fs makes and uses special meta information which is "x-amz-meta-***" and "Content-Type" as HTTP header.
The s3fs sets and uses HTTP headers for the directory object, those headers are listed below.
Content-Type: application/x-directory
x-amz-meta-mode: <mode>
x-amz-meta-uid: <UID>
x-amz-meta-gid <GID>
x-amz-meta-mtime: <unix time of modified file>
Other S3 client application builds the directory object without attributes which is needed by the s3fs.
When the "ls" command is run on the s3fs-fuse file system which has directories/files made by other S3 clients, this result is shown below.
d--------- 1 root root 0 Feb 27 11:21 dir
---------- 1 root root 1024 Mar 14 02:15 file
Because the objects don't have meta information("x-amz-meta-mode"), it means mode=0000.
In this case, the directory object is shown only "d", because the s3fs determines the object as a directory when the object is the name with "/" or has "Content-type: application/x-directory" header.
(The s3fs sets "Content-Type: application/x-directory" to the directory object, but other S3 clients set "binary/octet-stream".)
In that result, nobody without root is allowed to operate the object.
The owner and group are "root"(UID=0) because the object doesn't have "x-amz-meta-uid/gid".
If the object doesn't have "x-amz-meta-mtime", the s3fs uses "Last-Modified" HTTP header.
Therefore the object's mtime is "Last-Modified" value.(This logic is same as old version)
It has been already explained, if you need to change the object attributes, you can do it by manually operation or mergedir.sh.
* Example of the compatibility with s3cmd etc
** Case A) Only "dir/file" object
One of case, there is only "dir/file" object without "dir/" object, that object is made by s3cmd or etc.
In this case, the response of REST API(list bucket) with "delimiter=/" parameter has "CommonPrefixes", and the "dir/" is listed in "CommonPrefixes/Prefix", but the "dir/" object is not real object.
The s3fs needs to determine this object as directory, however there is no real directory object("dir" or "dir/").
But both new s3fs and old one does NOT understand this "dir/" in "CommonPrefixes", because the s3fs fails to get meta information from "dir" or "dir/".
On this case, the result of "ls" command is shown below.
??????????? ? ? ? ? ? dir
This "dir" is not operated by anyone and any process, because the s3fs does not understand this object permission.
And "dir/file" object can not be shown and operated too.
Some other S3 clients(tntDrive/Gladinet/etc) can not understand this object as same as the s3fs.
If you need to operate "dir/file" object, you need to make the "dir/" object as a directory.
To make the "dir/" directory object, you need to do below.
Because there is already the "dir" object which is not real object, you can not make "dir/" directory.
(s3cmd does not make "dir/" object because the object name has "/".).
You should make another name directory(ex: "dir2/"), and move the "dir/file" objects to in new directory.
Last, you can rename the directory name from "dir2/" to "dir/".
** Case B) Both "dir" and "dir/file" object
This case is that there are "dir" and "dir/file" objects which were made by s3cmd/etc.
s3cmd and s3fs understand the "dir" object as normal(file) object because this object does not have meta information and a name with "/".
But the result of REST API(list bucket) has "dir/" name in "CommonPrefixes/Prefix".
The s3fs checks "dir/" and "dir" as a directory, but the "dir" object is not directory object.
(Because the new s3fs need to compatible old version, the s3fs checks a directory object in order of "dir/", "dir")
In this case, the result of "ls" command is shown below.
---------- 1 root root 0 Feb 27 02:48 dir
As a result, the "dir/file" can not be shown and operated because the "dir" object is a file.
If you determine the "dir" as a directory, you need to add mete information to the "dir" object by s3cmd.
** Case C) Both "dir" and "dir/" object
Last case is that there are "dir" and "dir/" objects which were made by other S3 clients.
(example: At first you upload a object "dir/" as a directory by new 3sfs, and you upload a object "dir" by s3cmd.)
New s3fs determines "dir/" as a directory, because the s3fs searches in oder of "dir/", "dir".
As a result, the "dir" object can not be shown and operated.
** Compatibility between S3 clients
Both new and old s3fs do not understand both "dir" and "dir/" at the same time, tntDrive and Galdinet are same as the s3fs.
If there are "dir/" and "dir" objects, the s3fs gives priority to "dir/".
But s3cmd and DragonDisk understand both objects.
git-svn-id: http://s3fs.googlecode.com/svn/trunk@392 df820570-a93a-0410-bd06-b72b767a4274
1) Lifetime for the stats cache
Added the new option "stat_cache_expire".
This option which is specified by seconds means the lifetime for each stats cache entry.
If this option is not specified, the stats cache is kept in s3fs process until the stats cache grown to maximum size. (default)
If this option is specified, the stats cache entry is out from the memory when the entry expires time.
2) Enable file permission
s3fs before 1.62 did not consider the file access permission.
s3fs after this version can consider it.
For access permission, the s3fs_getattr() function was divided into sub function which can check the file access permission.
It is like access() function.
And the function calling the s3fs_getattr() calls this new sub function instead of s3fs_getattr().
Last the s3fs_opendir() function which is called by FUSE was added for checking directory access permission when listing the files in directory.
3) UID/GUID
When a file or a directory was created, the s3fs could not set the UID/GID as the user who executed a command.
(Almost the UID/GID are root, because the s3fs run by root.)
After this version, the s3fs set correct UID/GID as the user who executes the commond.
4) About the mtime
If the object does not have "x-amz-meta-mtime" meta, the s3fs uses the "Last-Modified" header instead of it.
But the s3fs had a bug in this code, and this version fixed this bug.
When user modified the file, the s3fs did not update the mtime of the file.
This version fixed this bug.
In the get_local_fd() function, the local file's mtime was changed only when s3fs run with "use_cache" option.
This version always updates the mtime whether the local cache file is used or not.
And s3fs_flush ( ) function set the mtime of local cache file from S3 object mtime, but it was wrong .
This version is that the s3fs_flush ( ) changes the mtime of S3 object from the local cache file or the tmpfile .
The s3fs cuts some requests, because the s3fs can always check mtime whether the s3fs uses or does not use the local cache file.
5) A case of no "x-amz-meta-mode"
If the object did not have "x-amz-meta-mtime" mete, the s3fs recognized the file as not regular file.
After this version, the s3fs recognizes the file as regular file.
6) "." and ".." directory
The s3fs_readdir() did not return "X" and "XX" directory name.
After this version, the s3fs is changed that it returns "X" and "XX".
Example, the result of "ls" lists "X" and "XX" directory.
7) Fixed a bug
The insert_object() had a bug, and it is fixed.
git-svn-id: http://s3fs.googlecode.com/svn/trunk@390 df820570-a93a-0410-bd06-b72b767a4274
2) Added a static table for file discriptors and paths which is keeped until closing.
The s3fs_xxxx function called by FUSE is able to use a file disctiptor which is already opened by s3fs_open function.
3) The mknod is changed that it always return error, because it does not work through a network.
4) The symbolic file attribute changed to S_IFLNK | S_IRWXU | S_IRWXG | S_IRWXO.
5) Fixed the truncate function to work.
6) The mkdir and clone_directory_object functions are simplified and are changed to use the common create_directory_object function for these.
For fixed a bug that the directory's PID/UID/mode are changed when these are renamed.
7) The get_object_name function is changed to check a object finely.
8) The s3fs_check_service function is changed for "301" response code.
9) Added the noxmlns option for a case of the response without xmlns field. (for the storage compatible with S3)
10) Added the nocopyapi option for the storage compatible with S3 without copy API.
* Comments
No.9 and No.10 are for the storage compatible with AWS S3.
Both option are unnecessary options for AWS S3.
In future, for the s3fs's promotion and possibility I would like to add new function.
git-svn-id: http://s3fs.googlecode.com/svn/trunk@384 df820570-a93a-0410-bd06-b72b767a4274
Modified rename_object and put_headers to handle objects larger than
5GB. Files larger than 5GB are required to use the multi interface.
git-svn-id: http://s3fs.googlecode.com/svn/trunk@363 df820570-a93a-0410-bd06-b72b767a4274
complete s3fs_readdir() refactor
- multi interface now batches HTTP requests
- proper HTTP KeepAlive sessions are back! (CURLOPT_FORBID_REUSE is no longer required)
- use xpath to quickly grab xml nodes
- lots of cleanup
- fixes some strange stat cache behavior
- huge readdir performance benefits (8-14x in my case) on large directories
git-svn-id: http://s3fs.googlecode.com/svn/trunk@348 df820570-a93a-0410-bd06-b72b767a4274
- fixed a bug in the file cache, it was attempting to set the mtime
on symlinks
- general code cleanup; moved some string functions to string_util.cpp
git-svn-id: http://s3fs.googlecode.com/svn/trunk@345 df820570-a93a-0410-bd06-b72b767a4274
- s3fs_flush() now checks to see whether the file on the remote end is the same as the local copy.
- md5sum() now requires a file descriptor instead of a path.
git-svn-id: http://s3fs.googlecode.com/svn/trunk@322 df820570-a93a-0410-bd06-b72b767a4274
Functional changes are limited to the multipart upload process. Each uploaded part is now verified against a local md5sum.
git-svn-id: http://s3fs.googlecode.com/svn/trunk@318 df820570-a93a-0410-bd06-b72b767a4274
Beginning of s3fs "utility" mode - initially -u option
just reports in progress multipart uploads for the
bucket. Eventually this mode could be used for
other S3 tasks not accessible through typical
file system operations
For multipart upload, use safer mkstemp() instead
of tmpnam() for temporary file
Increased the curl connect and readwrite timeouts
to 10 and 30 seconds respectively.
Autodetect when a big file is being uploaded,
increase the readwrite timeout to 120 seconds. This
was found through experimentation. When uploading
a big file, it is suspected that time is needed
for S3 to assemble the file before it is available
for access. It was found that when a large file
was uploaded via rsync, the final mtime and
chmod modifications were timing out, even though
the upload itself was successful.
Multipart upload is ready for use. A couple of
error checks are still needed in the function and
some cleanup. Need some feedback on how it
is working though.
git-svn-id: http://s3fs.googlecode.com/svn/trunk@298 df820570-a93a-0410-bd06-b72b767a4274
Check issue #142 for details
Code is operational, but not quite ready for
prime time -- needs some clean up
git-svn-id: http://s3fs.googlecode.com/svn/trunk@297 df820570-a93a-0410-bd06-b72b767a4274
refactoring easier and the code easier to understand (for me anyway)
Opened up the VERIFY macro so that memory cleanup can be done
before returning from a function.
Make the file descriptor function calls a bit more robust,
check the return codes.
Current code tested on Debian sid, CentOS (with FUSE 2.84) and Ubuntu 10.10
git-svn-id: http://s3fs.googlecode.com/svn/trunk@286 df820570-a93a-0410-bd06-b72b767a4274
No tarball until further testing on other platforms.
Extensively tested on Debian sid 64bit
Resolves issue #104
git-svn-id: http://s3fs.googlecode.com/svn/trunk@285 df820570-a93a-0410-bd06-b72b767a4274
expose the curl compiled with openssl vs. nss issue.
If the issue is seen, emit an informational message
and give the user an option to over-ride checking of
the hostname -- it's recommended not to use bucket
names with periods and https
As implied, added an option ssl_verify_hostname=[0|1]
Tested on fedora 14. Will check on Ubuntu/Debian/CentOS
after check in.
Resolves issue #128
git-svn-id: http://s3fs.googlecode.com/svn/trunk@270 df820570-a93a-0410-bd06-b72b767a4274
trapping the CURLE_SSL_CACERT error and trying to automatically
correct it.
Tested on CentOS 5.5 and ensured that Debian/Ubuntu doesn't break
because of this. This only applies to the https usage.
git-svn-id: http://s3fs.googlecode.com/svn/trunk@269 df820570-a93a-0410-bd06-b72b767a4274
directory from the trunk directory.
First do a svn cp of all of the source up to
trunk. This is supposed to preserve change
history -- we'll see.
The source remains untouched until this gets
worked out.
Also in preparation of bringing in the source
collateral for the debian package into the
repository. I expect that the top level will
look like this:
svn/
s3fs/
trunk/
tags/
branches/
dpkg/
trunk/
tags/
branches/
So far that's how it is looking. I'll be
very careful to ensure integrity of the data.
As a result this may be a multistep process.
git-svn-id: http://s3fs.googlecode.com/svn/trunk@236 df820570-a93a-0410-bd06-b72b767a4274
Andrew Gaul
Takeshi Nakatani
ggtakec@gmail.com
ben.lemasurier@gmail.com
mooredan@suncup.net
apetresc