From: Niels de Vos <address@hidden>
To:
Steve Bakke <address@hidden>
Cc: "address@hidden" <address@hidden>
Sent: Tuesday, September 18, 2012 3:16 AM
Subject: Re: [Gluster-devel] Fuse mount option: enable-ino32
On 09/18/2012 01:18 AM, Anand Avati wrote:
> The right place to do it is in fuse_readdir_cbk, based on
> priv->enable_ino32 (see c13823bd16b26bc471d3efb15f63b76fbfdf0309)
Yeah, it seems that fuse_readdir_cbk uses d_ino directly. Could you try the
attached patch? I probably won't have the time to test this in the next days,
but if you can confirm this works for you I can post it for review.
> On Mon, Sep 17, 2012 at 4:13 PM, Steve Bakke<
address@hidden> wrote:
>> Found a temporary solution. We patched
>> xlators/protocol/client/src/client-helpers.c ~ line 159
>> was: entry->d_ino = trav->d_ino
>> now: entry->d_ino = GF_FUSE_SQUASH_INO(trav->d_info);
This will work, but does not take the mount-option in consideration. When you
test the attached patch, please revert the above change.
Thanks,
Niels
>> All the 32 bit apps that did not work before, now work. Need both the
>> original enable-ino32 patch and this one.
>>
>> Not sure what the actual patch should be, but this does work at least for
>> our running only 32 bit apps on a Gluster mount.
>>
>> Steve
>>
>> ------------------------------
>> *From:* Steve Bakke<
address@hidden>
>> *To:* "
address@hidden"<
address@hidden>
>> *Sent:* Monday, September 17, 2012 11:42 AM
>> *Subject:* Re: [Gluster-devel] Fuse mount option: enable-ino32
>>
>> Yes we are running the patched version on client. We modified the print
>> line in our test program to display the inode values and compiled the
>> program for 64 bit mode, so it will run.
>> printf("inode:%016lx file:%s\n", dirEntry.d_ino, dirEntry.d_name );
>>
>>
>> Also to make sure client was working correctly we added logging calls to
>> gf_fuse_stat2attr() in fuse-helpers.c,
at line 444 added:
>>
>> fprintf( mylogfile, "ino32: %i st:%lx fuse:%lx \n", enable_ino32,
>> st->ia_ino, fa->ino );
>>
>>
>>
>> When the test program was run, it printed this out:
>>
>> inode:a2bdc7aeaee47329 file:steve.txt
>> inode:821222a0edf18a37 file:george.txt
>> inode:0000000000000001 file:.
>> inode:a7f29365717fb464 file:testnfs
>> inode:ffffffffffffffff file:..
>> inode:9aa3f33d3c366b4e file:fred.txt
>>
>>
>> The log file contained:
>>
>> ino32: 1 st:1 fuse:1
>> ino32: 1 st:1 fuse:1
>> ino32: 1 st:1 fuse:1
>> ino32: 1 st:a2bdc7aeaee47329 fuse:c59b487
>> ino32: 1 st:821222a0edf18a37 fuse:6fe3a897
>> ino32: 1 st:a7f29365717fb464 fuse:d68d2701
>> ino32: 1 st:9aa3f33d3c366b4e fuse:a6959873
>> ino32:
1 st:1 fuse:1
>> ino32: 1 st:1 fuse:1
>> ino32: 1 st:1 fuse:1
>> ino32: 1 st:1 fuse:1
>> ino32: 1 st:1 fuse:1
>>
>>
>> As the inode values return to the test program are> 32 bits, seems like
>> there is some path that the inodes are not getting adjusted?
>>
>> When we compile the test program in 32 bit mode, we get error 75, which is
>> EOVERFLOW. Looking at the 32bit Getdents (called by readdir) EOVERFLOW is
>> returned when inode values are> 32 bits.
>>
>> Steve
>>
>>
>> ------------------------------
>> *From:* Niels de Vos<
address@hidden>
>> *To:* Steve Bakke<
address@hidden>
>>
*Cc:* "
address@hidden"<
address@hidden>
>> *Sent:* Monday, September 17, 2012 3:49 AM
>> *Subject:* Re: [Gluster-devel] Fuse mount option: enable-ino32
>>
>> On 09/15/2012 11:15 PM, Steve Bakke wrote:
>>> All,
>>>
>>> Just compiled and installed the Sept 15 git which includes the fuse
>> mount option enable-ino32 patch.
>>>
>>> Then on the server:
>>> gluster volume fred set enable-ino32 on
>>> Note: volume info prints out nfs.enable-ino32: on, which I guess is ok???
>>
>> The volume option is for the NFS-server only. It does not affect the
>> GlusterFS-native mount, therefore there is no need to set
this option for
>> the
>> volume.
>>
>>> On client:
>>> mount -t glusterfs -o enable-ino32 1.2.3.4:/fred mntfs
>>
>> This is the correct usage.
>>
>>> Mounts, but still returns wrong entries, using the test program below.
>> This program works fine with using nfs mount, and with nfs.enable-ino32 on
>> server. However for other reasons must use native client.
>>>
>>> Do I need to do something else??
>>
>> Are you using the version with the patch on the client? The conversion of
>> the
>> 64-bit to 32-bit inodes is done client-side, not server-side. This makes
>> it
>> possible for clients to choose if they want/need 32- or 64-bit inodes.
>>
>> Cheers,
>> Niels
>>
>>
>>>
Steve
>>>
>>>
>>>
>>> #include<stdlib.h>
>>> #include<stdio.h>
>>> #include<sys/types.h>
>>> #include<dirent.h>
>>>
>>> int main()
>>> {
>>> DIR* hDir = opendir(".");
>>> if (hDir==NULL)
>>> return 1;
>>>
>>> struct dirent dirEntry;
>>> struct dirent * pOut;
>>> int err = readdir_r(hDir,&dirEntry,&pOut);
>>> while(err==0&& pOut!=NULL)
>>> {
>>> printf("%s ", dirEntry.d_name);
>>> err = readdir_r(hDir,&dirEntry,&pOut);
>>> }
>>>
>>>
printf("\nLast Error: %d\n", err);
>>>
>>> return 0;
>>> }
>>>