Details
-
Bug
-
Resolution: Fixed
-
Critical
-
None
-
None
-
-
3
-
12838
Description
While compiling e2fsprogs against Lustre master branch on RHEL6.5/ppc64, I hit the following errors:
In file included from /root/lustre-release/build/lustre-2.5.56//lustre/include/lustre/lustre_user.h:51,
from /root/lustre-release/build/lustre-2.5.56//lustre/include/lustre/lustreapi.h:46,
from ../lib/ext2fs/lfsck.h:10,
from lfsck_common.c:29:
/root/lustre-release/build/lustre-2.5.56//libcfs/include/libcfs/posix/posix-types.h:48: error: conflicting types for 'umode_t'
/usr/include/asm/types.h:31: note: previous declaration of 'umode_t' was here
CC util.c
In file included from /root/lustre-release/build/lustre-2.5.56//lustre/include/lustre/lustre_user.h:51,
from /root/lustre-release/build/lustre-2.5.56//lustre/include/lustre/lustreapi.h:46,
from ../lib/ext2fs/lfsck.h:10,
from lfsck.c:80:
/root/lustre-release/build/lustre-2.5.56//libcfs/include/libcfs/posix/posix-types.h:48: error: conflicting types for 'umode_t'
/usr/include/asm/types.h:31: note: previous declaration of 'umode_t' was here
make[3]: *** [lfsck_common.o] Error 1
The build log "build_e2fsprogs.ppc64.log" is attached.
Attachments
Issue Links
- is related to
-
-
- Closed
-
Activity
By using the following temporary fix suggested by nasf, the number of failed tests was reduced from 37 to 3 (f_eofblocks, f_pgsize_gt_blksize, f_dirdata):
diff --git a/lib/ext2fs/ext2_fs.h b/lib/ext2fs/ext2_fs.h
index a99ab37..95233ba 100644
--- a/lib/ext2fs/ext2_fs.h
+++ b/lib/ext2fs/ext2_fs.h
@@ -785,11 +785,16 @@ struct ext2_dir_entry {
* file_type field.
*/
struct ext2_dir_entry_2 {
- __u32 inode; /* Inode number */
- __u16 rec_len; /* Directory entry length */
- __u8 name_len; /* Name length */
- __u8 file_type;
- char name[EXT2_NAME_LEN]; /* File name */
+ __u32 inode; /* Inode number */
+ __u16 rec_len; /* Directory entry length */
+#ifndef WORDS_BIGENDIAN
+ __u8 name_len; /* Name length */
+ __u8 file_type;
+#else
+ __u8 file_type;
+ __u8 name_len; /* Name length */
+#endif
+ char name[EXT2_NAME_LEN]; /* File name */
};
After temporarily disabling the 3 failed tests, I successfully built e2fsprogs on the ppc64 node and got the rpm packages.
I'll see what errors I'll hit while testing Lustre server codes on the ppc64 node. In the meantime, nasf will continue digging into the failures and find out more proper way to resolve those failures. Thanks, nasf!
It should be related with bytes order in the structure of "ext2_dirent_entry" and "ext2_dirent_entry_2" on the PPC64 machine.
struct ext4_dir_entry {
__le32 inode; /* Inode number */
__le16 rec_len; /* Directory entry length */
__le16 name_len; /* Name length */
char name[EXT4_NAME_LEN]; /* File name */
};
/*
* The new version of the directory entry. Since EXT4 structures are
* stored in intel byte order, and the name_len field could never be
* bigger than 255 chars, it's safe to reclaim the extra byte for the
* file_type field.
*/
struct ext4_dir_entry_2 {
__le32 inode; /* Inode number */
__le16 rec_len; /* Directory entry length */
__u8 name_len; /* Name length */
__u8 file_type;
char name[EXT4_NAME_LEN]; /* File name */
};
In fact, the ext4_dir_entry_2::name_len and ext4_dir_entry_2::file_type should be exchanged on PPC64 as following:
struct ext2_dir_entry_2 {
__le32 inode; /* Inode number */
__le16 rec_len; /* Directory entry length */
#ifndef WORDS_BIGENDIAN
__u8 name_len; /* Name length */
__u8 file_type;
#else
__u8 file_type;
__u8 name_len; /* Name length */
#endif
char name[EXT2_NAME_LEN]; /* File name */
};
Then we can access the ext2_dir_entry::name_len via ext2_dir_entry_2::name_len.
In fact, because the CPU bytes-order for PPC64 and the on-disk bytes-order are different, any value (not only ext2_dir_dirent, but also others, such as inode, super_block, and so on) which is larger than 1-byte and loaded from disk (write or disk) should be converted via lexx_to_cpu (or cpu_to_lexx). Unfortunately, we missed to do that at a lot of points in current implantation.
On the other hand, in current implementation, we assume the ext2_dir_entry_2::name_len is lower than ext2_dir_entry_2::file_type, so there are some logic like the following:
int filetype = dirent->name_len >> 8;
...
dirent->name_len = ((dirent->name_len & 0xFF) | (dirdata | should_be) << 8);
...
To make the e2fsck workable on PPC64 correctly, all related codes should be well checked and re-considere. In theory, it is not complex, but it involves a lot of trivial things and I am not sure whether there will be other data structures to be handled. We must check carefully one by one. So it is not a simple work.
I;ve added Fanyong to the list because he can provide expertise in this topic.
It is my guess that the base e2fsprogs (before the "Lustre version" patch will not have this problem.
Yes, Andreas, I verified this.
Here are the commits on e2fsprogs master-lustre branch: http://git.whamcloud.com/?p=tools/e2fsprogs.git;a=shortlog;h=refs/heads/master-lustre
Since commit f6632828e4845041bce12f79d8f3aae095534bc6 (e2fsck: handle preallocation for large PAGE_SIZE), "make check" started failing on ppc64 node. And before that commit, "make check" passed.
Here are the details about which test failures on ppc64 node are introduced by which commits:
| Failed tests | Introduced by |
|---|---|
| f_eofblocks | e2fsck: handle preallocation for large PAGE_SIZE |
| f_pgsize_gt_blksize | tests: PAGE_SIZE larger than blocksize with hole |
| f_badsymlinks f_dup_de2 f_dup_de f_ext_journal f_h_badnode f_h_badroot f_h_reindex f_jchksum_remount f_rehash_dir f_uninit_bad_free_inodes f_uninit_blk_used_not_set f_uninit_set_inode_not_set f_zero_group f_zero_super | e2fsck: add support for dir_data feature |
| f_dirdata | tests: add basic tests for dir_data feature |
| f_dirdata_optimize f_dir_optimize | |
| f_random_corruption | debugfs: dump "fid" and "lma" xattrs on inode stat |
| s_basic_scan | tests: add basic test case for e2scan |
| d_loaddump d_special_files f_badroot f_badtable f_dup4 f_dupdot f_dup_resize f_expand f_illitable f_lpf f_lpffile f_noroot f_recnect_bad f_reconnect m_no_opt t_quota_1on |
Could you and Jinshan please give me some instructions about how to make Lustre e2fsprogs codes pass building on ppc64 node without affecting the testing for LU-4665? Should I look into the above test failures and resolve them? Thanks a lot in advance.
Note that we have probably never compiled the Lustre-patches e2fsprogs code on PPC. When I was suggesting to bisect the code, I was thinking about the patches that we apply on top of the base e2fsprogs. It is my guess that the base e2fsprogs (before the "Lustre version" patch will not have this problem.
I can work with Yujian on this - do you have a rough idea about what was the latest working version?
It is possible that this error is caused by the dirdata patch. I would suggest to use "git bisect" to try and isolate it quickly. Please ping me in Skype if you haven't used this before and want some help.
Hi Andreas & Alex,
Do you have any idea for the error mentioned by Yujian?
Jinshan
With patch #9437, e2fsprogs passed building on ppc64 node.
However, it failed at "Running e2fsprogs test suite..." stage:
Running e2fsprogs test suite...
/bin/cp ./mke2fs.conf.in mke2fs.conf
Creating test_one script...
Creating test_script...
cmp: d_loaddump.ver.tmp: No such file or directory
d_loaddump: debugfs load/dump test: failed
— d_loaddump/expect 2014-02-27 09:06:42.000000000 +0000
+++ d_loaddump.log 2014-02-28 14:45:47.251311888 +0000
@@ -7,12 +7,26 @@
e2fsck -yf -N test_filesys
Pass 1: Checking inodes, blocks, and sizes
Pass 2: Checking directory structure
+Directory inode 2, block #0, offset 24: directory corrupted
+Salvage? yes
+
Pass 3: Checking directory connectivity
+Unconnected directory inode 11 (/???)
+Connect to /lost+found? yes
+
+/lost+found not found. Create? yes
+
+Pass 3A: Optimizing directories
Pass 4: Checking reference counts
+Inode 11 ref count is 3, should be 2. Fix? yes
+
Pass 5: Checking group summary information
-test_filesys: 12/64 files (0.0% non-contiguous), 158/512 blocks
-Exit status is 0
+
+test_filesys: ***** FILE SYSTEM WAS MODIFIED *****
+test_filesys: 13/64 files (0.0% non-contiguous), 148/512 blocks
+Exit status is 1
debugfs -R ''dump test_data d_loaddump.ver.tmp'' test.img
+test_data: EXT2 directory corrupted
Exit status is 0
cmp d_loaddump.tmp d_loaddump.ver.tmp
-Exit status is 0
+Exit status is 2
-----------8<-----------
118 tests succeeded 37 tests failed
Tests failed: d_loaddump d_special_files f_badroot f_badsymlinks f_badtable f_dir_optimize f_dirdata f_dirdata_optimize f_dup4 f_dup_de f_dup_de2 f_dup_resize f_dupdot f_eofblocks f_expand f_ext_journal f_h_badnode f_h_badroot f_h_reindex f_illitable f_jchksum_remount f_lpf f_lpffile f_noroot f_pgsize_gt_blksize f_random_corruption f_recnect_bad f_reconnect f_rehash_dir f_uninit_bad_free_inodes f_uninit_blk_used_not_set f_uninit_set_inode_not_set f_zero_group f_zero_super m_no_opt s_basic_scan t_quota_1on
make[2]: *** [test_post] Error 1
The log of "build_e2fsprogs_rpm.ppc64.log" is attached.
Looking at the e2fsprogs code, it seems that there is almost no code using ext2_dir_entry_2, it all uses ext2_dir_entry and then masks name_len and shifts down file_type. Unfortunately, in the e2fsprogs code, the in-memory representation of the directory entry is often already swabbed, so the change to ext2_dir_entry_2 would break for entries that are already swabbed.
I think it would be better to update the Lustre patches to use ext2_dir_entry and use the mask-and-shift method used by the rest of the code. Since I'm already updating the e2fsprogs patches to the latest maint release, I will do this part. I'm not sure what is wrong with the PAGE_SIZE patches or others since I don't know why they are failing.