From: "Darrick J. Wong" <darrick.wong@oracle.com>
To: tytso@mit.edu, darrick.wong@oracle.com
Cc: linux-ext4@vger.kernel.org
Subject: [PATCH 43/54] copy-in: for files, only iterate file blocks that are mapped
Date: Mon, 26 Jan 2015 23:40:15 -0800 [thread overview]
Message-ID: <20150127074015.13308.22535.stgit@birch.djwong.org> (raw)
In-Reply-To: <20150127073533.13308.44994.stgit@birch.djwong.org>
Rewrite the file copy-in algorithm to detect smaller holes in the
files we're copying in. Use SEEK_DATA/SEEK_HOLE/FIEMAP when available
to skip known empty parts. This fixes the particular bug where zeroed
blocks on a system with 64k pages are needlessly copied into a
4k-block filesystem. It also saves time by skipping parts we know to
be zeroed.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
---
misc/create_inode.c | 280 ++++++++++++++++++++++++++++++++++++---------------
1 file changed, 196 insertions(+), 84 deletions(-)
diff --git a/misc/create_inode.c b/misc/create_inode.c
index 3bc0515..8ab546e 100644
--- a/misc/create_inode.c
+++ b/misc/create_inode.c
@@ -9,12 +9,22 @@
* %End-Header%
*/
+#define _FILE_OFFSET_BITS 64
+#define _LARGEFILE64_SOURCE 1
+#define _GNU_SOURCE 1
+
+#include "config.h"
#include <time.h>
+#include <sys/types.h>
#include <unistd.h>
#include <limits.h> /* for PATH_MAX */
#ifdef HAVE_ATTR_XATTR_H
#include <attr/xattr.h>
#endif
+#include <sys/ioctl.h>
+#include <ext2fs/ext2fs.h>
+#include <ext2fs/ext2_types.h>
+#include <ext2fs/fiemap.h>
#include "create_inode.h"
#include "nls-enable.h"
@@ -28,14 +38,7 @@
#endif
/* 64KiB is the minimium blksize to best minimize system call overhead. */
-#ifndef IO_BUFSIZE
-#define IO_BUFSIZE 64*1024
-#endif
-
-/* Block size for `st_blocks' */
-#ifndef S_BLKSIZE
-#define S_BLKSIZE 512
-#endif
+#define COPY_FILE_BUFLEN 65536
static int ext2_file_type(unsigned int mode)
{
@@ -139,10 +142,9 @@ static errcode_t set_inode_xattr(ext2_filsys fs, ext2_ino_t ino, const char *fil
{
#ifdef HAVE_LLISTXATTR
errcode_t retval, close_retval;
- struct ext2_inode inode;
struct ext2_xattr_handle *handle;
ssize_t size, value_size;
- char *list;
+ char *list = NULL;
int i;
size = llistxattr(filename, NULL, 0);
@@ -382,82 +384,202 @@ try_again:
return retval;
}
-static errcode_t copy_file(ext2_filsys fs, int fd, ext2_ino_t newfile,
- int bufsize, int make_holes)
+#if !defined HAVE_PREAD64 && !defined HAVE_PREAD
+static ssize_t my_pread(int fd, const void *buf, size_t count, off_t offset)
{
- ext2_file_t e2_file;
- errcode_t retval, close_ret;
- int got;
- unsigned int written;
- char *buf;
- char *ptr;
- char *zero_buf;
- int cmp;
-
- retval = ext2fs_file_open(fs, newfile,
- EXT2_FILE_WRITE, &e2_file);
- if (retval)
- return retval;
-
- retval = ext2fs_get_mem(bufsize, &buf);
- if (retval) {
- com_err("copy_file", retval, "can't allocate buffer\n");
- goto out_close;
- }
+ if (lseek(fd, offset, SEEK_SET) < 0)
+ return 0;
- /* This is used for checking whether the whole block is zero */
- retval = ext2fs_get_memzero(bufsize, &zero_buf);
- if (retval) {
- com_err("copy_file", retval, "can't allocate zero buffer\n");
- goto out_free_buf;
- }
+ return read(fd, buf, count);
+}
+#endif /* !defined HAVE_PREAD64 && !defined HAVE_PREAD */
- while (1) {
- got = read(fd, buf, bufsize);
- if (got == 0)
- break;
+static errcode_t copy_file_range(ext2_filsys fs, int fd, ext2_file_t e2_file,
+ off_t start, off_t end, char *buf,
+ char *zerobuf)
+{
+ off_t off, bpos;
+ ssize_t got, blen;
+ unsigned int written;
+ char *ptr;
+ errcode_t err = 0;
+
+ for (off = start; off < end; off += COPY_FILE_BUFLEN) {
+#ifdef HAVE_PREAD64
+ got = pread64(fd, buf, COPY_FILE_BUFLEN, off);
+#elif HAVE_PREAD
+ got = pread(fd, buf, COPY_FILE_BUFLEN, off);
+#else
+ got = my_pread(fd, buf, COPY_FILE_BUFLEN, off);
+#endif
if (got < 0) {
- retval = errno;
+ err = errno;
goto fail;
}
- ptr = buf;
-
- /* Sparse copy */
- if (make_holes) {
- /* Check whether all is zero */
- cmp = memcmp(ptr, zero_buf, got);
- if (cmp == 0) {
- /* The whole block is zero, make a hole */
- retval = ext2fs_file_lseek(e2_file, got,
- EXT2_SEEK_CUR,
- NULL);
- if (retval)
+ for (bpos = 0, ptr = buf; bpos < got; bpos += fs->blocksize) {
+ blen = fs->blocksize;
+ if (blen > got - bpos)
+ blen = got - bpos;
+ if (memcmp(ptr, zerobuf, blen) == 0) {
+ ptr += blen;
+ continue;
+ }
+ err = ext2fs_file_lseek(e2_file, off + bpos,
+ EXT2_SEEK_SET, NULL);
+ if (err)
+ goto fail;
+ while (blen > 0) {
+ err = ext2fs_file_write(e2_file, ptr, blen,
+ &written);
+ if (err)
+ goto fail;
+ if (written == 0) {
+ err = EIO;
goto fail;
- got = 0;
+ }
+ blen -= written;
+ ptr += written;
}
}
+ }
+fail:
+ return err;
+}
- /* Normal copy */
- while (got > 0) {
- retval = ext2fs_file_write(e2_file, ptr,
- got, &written);
- if (retval)
- goto fail;
-
- got -= written;
- ptr += written;
+static errcode_t try_lseek_copy(ext2_filsys fs, int fd, struct stat *statbuf,
+ ext2_file_t e2_file, char *buf, char *zerobuf)
+{
+#if defined(SEEK_DATA) && defined(SEEK_HOLE)
+ off_t data = 0, hole;
+ off_t data_blk, hole_blk;
+ errcode_t err;
+
+ /* Try to use SEEK_DATA and SEEK_HOLE */
+ while (data < statbuf->st_size) {
+ data = lseek(fd, data, SEEK_DATA);
+ if (data < 0) {
+ if (errno == ENXIO)
+ break;
+ return EXT2_ET_UNIMPLEMENTED;
}
+ hole = lseek(fd, data, SEEK_HOLE);
+ if (hole < 0)
+ return EXT2_ET_UNIMPLEMENTED;
+
+ data_blk = data & ~(fs->blocksize - 1);
+ hole_blk = (hole + (fs->blocksize - 1)) & ~(fs->blocksize - 1);
+ err = copy_file_range(fs, fd, e2_file, data_blk, hole_blk, buf,
+ zerobuf);
+ if (err)
+ return err;
+
+ data = hole;
}
-fail:
- ext2fs_free_mem(&zero_buf);
-out_free_buf:
+ return err;
+#else
+ return EXT2_ET_UNIMPLEMENTED;
+#endif /* SEEK_DATA and SEEK_HOLE */
+}
+
+static errcode_t try_fiemap_copy(ext2_filsys fs, int fd, ext2_file_t e2_file,
+ char *buf, char *zerobuf)
+{
+#if defined(FS_IOC_FIEMAP)
+#define EXTENT_MAX_COUNT 512
+ struct fiemap *fiemap_buf;
+ struct fiemap_extent *ext_buf, *ext;
+ int ext_buf_size, fie_buf_size;
+ off_t pos = 0;
+ unsigned int i;
+ errcode_t err;
+
+ ext_buf_size = EXTENT_MAX_COUNT * sizeof(struct fiemap_extent);
+ fie_buf_size = sizeof(struct fiemap) + ext_buf_size;
+
+ err = ext2fs_get_memzero(fie_buf_size, &fiemap_buf);
+ if (err)
+ return err;
+
+ ext_buf = fiemap_buf->fm_extents;
+ memset(fiemap_buf, 0, fie_buf_size);
+ fiemap_buf->fm_length = FIEMAP_MAX_OFFSET;
+ fiemap_buf->fm_flags |= FIEMAP_FLAG_SYNC;
+ fiemap_buf->fm_extent_count = EXTENT_MAX_COUNT;
+
+ do {
+ fiemap_buf->fm_start = pos;
+ memset(ext_buf, 0, ext_buf_size);
+ err = ioctl(fd, FS_IOC_FIEMAP, fiemap_buf);
+ if (err < 0 && (errno == EOPNOTSUPP || errno == ENOTTY)) {
+ err = EXT2_ET_UNIMPLEMENTED;
+ goto out;
+ } else if (err < 0 || fiemap_buf->fm_mapped_extents == 0) {
+ err = errno;
+ goto out;
+ }
+ for (i = 0, ext = ext_buf; i < fiemap_buf->fm_mapped_extents;
+ i++, ext++) {
+ err = copy_file_range(fs, fd, e2_file, ext->fe_logical,
+ ext->fe_logical + ext->fe_length,
+ buf, zerobuf);
+ if (err)
+ goto out;
+ }
+
+ ext--;
+ /* Record file's logical offset this time */
+ pos = ext->fe_logical + ext->fe_length;
+ /*
+ * If fm_extents array has been filled and
+ * there are extents left, continue to cycle.
+ */
+ } while (fiemap_buf->fm_mapped_extents == EXTENT_MAX_COUNT &&
+ !(ext->fe_flags & FIEMAP_EXTENT_LAST));
+out:
+ ext2fs_free_mem(&fiemap_buf);
+ return err;
+#else
+ return EXT2_ET_UNIMPLEMENTED;
+#endif /* FS_IOC_FIEMAP */
+}
+
+static errcode_t copy_file(ext2_filsys fs, int fd, struct stat *statbuf,
+ ext2_ino_t ino)
+{
+ ext2_file_t e2_file;
+ char *buf = NULL, *zerobuf = NULL;
+ errcode_t err, close_err;
+
+ err = ext2fs_file_open(fs, ino, EXT2_FILE_WRITE, &e2_file);
+ if (err)
+ return err;
+
+ err = ext2fs_get_mem(COPY_FILE_BUFLEN, &buf);
+ if (err)
+ goto out;
+
+ err = ext2fs_get_memzero(fs->blocksize, &zerobuf);
+ if (err)
+ goto out;
+
+ err = try_lseek_copy(fs, fd, statbuf, e2_file, buf, zerobuf);
+ if (err != EXT2_ET_UNIMPLEMENTED)
+ goto out;
+
+ err = try_fiemap_copy(fs, fd, e2_file, buf, zerobuf);
+ if (err != EXT2_ET_UNIMPLEMENTED)
+ goto out;
+
+ err = copy_file_range(fs, fd, e2_file, 0, statbuf->st_size, buf,
+ zerobuf);
+out:
+ ext2fs_free_mem(&zerobuf);
ext2fs_free_mem(&buf);
-out_close:
- close_ret = ext2fs_file_close(e2_file);
- if (retval == 0)
- retval = close_ret;
- return retval;
+ close_err = ext2fs_file_close(e2_file);
+ if (err == 0)
+ err = close_err;
+ return err;
}
static int is_hardlink(struct hdlinks_s *hdlinks, dev_t dev, ino_t ino)
@@ -481,8 +603,6 @@ errcode_t do_write_internal(ext2_filsys fs, ext2_ino_t cwd, const char *src,
ext2_ino_t newfile;
errcode_t retval;
struct ext2_inode inode;
- int bufsize = IO_BUFSIZE;
- int make_holes = 0;
fd = ext2fs_open_file(src, O_RDONLY, 0);
if (fd < 0) {
@@ -570,17 +690,9 @@ errcode_t do_write_internal(ext2_filsys fs, ext2_ino_t cwd, const char *src,
}
}
if (LINUX_S_ISREG(inode.i_mode)) {
- if (statbuf.st_blocks < statbuf.st_size / S_BLKSIZE) {
- make_holes = 1;
- /*
- * Use I/O blocksize as buffer size when
- * copying sparse files.
- */
- bufsize = statbuf.st_blksize;
- }
- retval = copy_file(fs, fd, newfile, bufsize, make_holes);
+ retval = copy_file(fs, fd, &statbuf, newfile);
if (retval)
- com_err("copy_file", retval, 0);
+ com_err("copy_file", retval, _("while copying %s"), src);
}
close(fd);
next prev parent reply other threads:[~2015-01-27 7:40 UTC|newest]
Thread overview: 88+ messages / expand[flat|nested] mbox.gz Atom feed top
2015-01-27 7:35 [PATCH 00/54] e2fsprogs January 2015 patchbomb Darrick J. Wong
2015-01-27 7:35 ` [PATCH 01/54] misc: fix minor testcase problems Darrick J. Wong
2015-01-27 15:55 ` Theodore Ts'o
2015-01-27 7:35 ` [PATCH 02/54] debugfs: document new commands Darrick J. Wong
2015-01-27 15:56 ` Theodore Ts'o
2015-01-27 7:35 ` [PATCH 03/54] debugfs: fix crash in ea_set argument handling Darrick J. Wong
2015-01-27 15:58 ` Theodore Ts'o
2015-01-27 7:35 ` [PATCH 04/54] libext2fs: initialize i_extra_isize when writing EAs Darrick J. Wong
2015-01-27 16:02 ` Theodore Ts'o
2015-01-27 7:36 ` [PATCH 05/54] libext2fs: avoid pointless EA block allocation Darrick J. Wong
2015-01-27 16:07 ` Theodore Ts'o
2015-01-27 19:26 ` Darrick J. Wong
2015-01-27 7:36 ` [PATCH 06/54] libext2fs: strengthen i_extra_isize checks when reading/writing xattrs Darrick J. Wong
2015-01-27 16:08 ` Theodore Ts'o
2015-01-27 7:36 ` [PATCH 07/54] libext2fs: fix tdb.c mmap leak Darrick J. Wong
2015-01-27 16:09 ` Theodore Ts'o
2015-01-27 7:36 ` [PATCH 08/54] resize2fs: fix regression test to not depend on ext4.ko being loaded Darrick J. Wong
2015-01-27 16:10 ` Theodore Ts'o
2015-01-27 7:36 ` [PATCH 09/54] tune2fs: disable csum verification before resizing inode Darrick J. Wong
2015-01-27 16:11 ` Theodore Ts'o
2015-01-27 7:36 ` [PATCH 10/54] tune2fs: abort when trying to enable/disable metadata_csum on mounted fs Darrick J. Wong
2015-01-27 16:26 ` Theodore Ts'o
2015-01-27 7:36 ` [PATCH 11/54] tune2fs: call out to resize2fs for 64bit conversion Darrick J. Wong
2015-01-27 16:31 ` Theodore Ts'o
2015-01-27 7:36 ` [PATCH 12/54] e2fsck: clear i_block[] when there are too many bad mappings on a special inode Darrick J. Wong
2015-01-27 16:32 ` Theodore Ts'o
2015-01-27 7:36 ` [PATCH 13/54] e2fsck: on read error, don't rewrite blocks past the end of the fs Darrick J. Wong
2015-01-27 17:35 ` Theodore Ts'o
2015-01-28 23:35 ` Darrick J. Wong
2015-01-27 7:37 ` [PATCH 14/54] e2fsck: fix the journal recreation message Darrick J. Wong
2015-01-27 18:02 ` Theodore Ts'o
2015-01-27 19:37 ` Darrick J. Wong
2015-01-27 7:37 ` [PATCH 15/54] e2fsck: handle multiple *ind block collisions with critical metadata Darrick J. Wong
2015-01-28 13:52 ` Theodore Ts'o
2015-01-27 7:37 ` [PATCH 16/54] e2fsck: decrement bad count _after_ remapping a duplicate block Darrick J. Wong
2015-01-28 13:58 ` Theodore Ts'o
2015-01-27 7:37 ` [PATCH 17/54] e2fsck: inspect inline dir data as two directory blocks Darrick J. Wong
2015-01-28 15:16 ` Theodore Ts'o
2015-01-27 7:37 ` [PATCH 18/54] e2fsck: improve the inline directory detector Darrick J. Wong
2015-01-28 16:38 ` Theodore Ts'o
2015-01-27 7:37 ` [PATCH 19/54] e2fsck: salvage under-sized dirents by removing them Darrick J. Wong
2015-02-16 15:40 ` Theodore Ts'o
2015-01-27 7:37 ` [PATCH 20/54] e2fsck: add a 'yes to all' response in interactive mode Darrick J. Wong
2015-03-29 2:54 ` Theodore Ts'o
2015-01-27 7:37 ` [PATCH 21/54] libext2fs: zero blocks via FALLOC_FL_ZERO_RANGE in ext2fs_zero_blocks Darrick J. Wong
2015-03-29 3:46 ` Theodore Ts'o
2015-01-27 7:37 ` [PATCH 22/54] libext2fs: ext2fs_new_block2() should call alloc_block hook Darrick J. Wong
2015-03-29 3:08 ` Theodore Ts'o
2015-01-27 7:38 ` [PATCH 23/54] libext2fs: Support readonly filesystem images Darrick J. Wong
2015-03-19 21:32 ` [PATCH v2 " Darrick J. Wong
2015-03-29 3:42 ` Theodore Ts'o
2015-01-27 7:38 ` [PATCH 24/54] libext2fs/e2fsck: provide routines to read-ahead metadata Darrick J. Wong
2015-01-27 7:38 ` [PATCH 25/54] e2fsck: read-ahead metadata during passes 1, 2, and 4 Darrick J. Wong
2015-01-27 7:38 ` [PATCH 26/54] e2fsck: track directories to be rehashed with a bitmap Darrick J. Wong
2015-01-27 7:38 ` [PATCH 27/54] e2fsck: rebuild sparse extent trees/convert non-extent ext3 files Darrick J. Wong
2015-03-19 21:42 ` [PATCH v4 " Darrick J. Wong
2015-01-27 7:38 ` [PATCH 28/54] tests: verify proper rebuilding of sparse extent trees and block map file conversion Darrick J. Wong
2015-01-27 7:38 ` [PATCH 29/54] undo-io: add new calls to and speed up the undo io manager Darrick J. Wong
2015-01-27 7:38 ` [PATCH 30/54] undo-io: be more flexible about setting block size Darrick J. Wong
2015-01-27 7:38 ` [PATCH 31/54] undo-io: use a bitmap to track what we've already written Darrick J. Wong
2015-01-27 7:39 ` [PATCH 32/54] e2undo: fix memory leaks and tweak the error messages somewhat Darrick J. Wong
2015-01-27 7:39 ` [PATCH 33/54] e2undo: ditch tdb file, write everything to a flat file Darrick J. Wong
2015-01-27 7:39 ` [PATCH 34/54] libext2fs: support atexit cleanups Darrick J. Wong
2015-01-27 7:39 ` [PATCH 35/54] e2fsck: optionally create an undo file Darrick J. Wong
2015-01-27 7:39 ` [PATCH 36/54] resize2fs: optionally create " Darrick J. Wong
2015-01-27 7:39 ` [PATCH 37/54] tune2fs: " Darrick J. Wong
2015-01-27 7:39 ` [PATCH 38/54] mke2fs: " Darrick J. Wong
2015-01-27 7:39 ` [PATCH 39/54] debugfs: " Darrick J. Wong
2015-01-27 7:39 ` [PATCH 40/54] tests: test undo file creation in e2fsck/resize2fs/tune2fs/mke2fs Darrick J. Wong
2015-01-27 7:40 ` [PATCH 41/54] tests: test various features of the new e2undo format Darrick J. Wong
2015-01-27 7:40 ` [PATCH 42/54] copy-in: create hardlinks with the correct directory filetype Darrick J. Wong
2015-01-27 7:40 ` Darrick J. Wong [this message]
2015-01-27 7:40 ` [PATCH 44/54] copyin: fix error handling Darrick J. Wong
2015-01-27 7:40 ` [PATCH 45/54] mke2fs: add simple tests and re-alphabetize mke2fs manpage options Darrick J. Wong
2015-01-27 7:40 ` [PATCH 46/54] contrib: script to create minified ext4 image from a directory Darrick J. Wong
2015-01-27 7:40 ` [PATCH 47/54] libext2fs: support allocating uninit blocks in bmap2() Darrick J. Wong
2015-01-27 7:40 ` [PATCH 48/54] libext2fs: find/alloc a range of empty blocks Darrick J. Wong
2015-01-27 7:40 ` [PATCH 49/54] libext2fs: add new hooks to support large allocations Darrick J. Wong
2015-01-27 7:41 ` [PATCH 50/54] libext2fs: implement fallocate Darrick J. Wong
2015-01-27 7:41 ` [PATCH 51/54] libext2fs: use fallocate for creating journals and hugefiles Darrick J. Wong
2015-01-27 7:41 ` [PATCH 52/54] debugfs: implement fallocate Darrick J. Wong
2015-01-27 7:41 ` [PATCH 53/54] tests: test debugfs punch command Darrick J. Wong
2015-03-19 21:44 ` [PATCH 55/54] e2fsck: actually fix inline_data flags problems when user says to do so Darrick J. Wong
2015-03-29 4:05 ` Theodore Ts'o
2015-03-19 21:45 ` [PATCH 56/54] libext2fs: zero hash in ibody extended attributes Darrick J. Wong
2015-03-29 4:13 ` Theodore Ts'o
2015-03-19 21:47 ` [PATCH 57/54] e2fsck: convert block-mapped files to extents on bigalloc fs Darrick J. Wong
2015-03-19 23:54 ` [PATCH 58/54] e2fsck: turn inline data symlink into a fast symlink when possible Darrick J. Wong
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20150127074015.13308.22535.stgit@birch.djwong.org \
--to=darrick.wong@oracle.com \
--cc=linux-ext4@vger.kernel.org \
--cc=tytso@mit.edu \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).