Re: 3TB disk hassles

[linux-os <linux-os@chaos.analogic.com>]

On Mon, 3 Jan 2005, Jeff V. Merkey wrote:

> Jeff V. Merkey wrote:
>
>> H. Peter Anvin wrote:
>> 
>>> Andries Brouwer wrote:
>>> 
>>>> 
>>>> Concerning one, it is a somewhat complicated format that takes over
>>>> your disk, rather inconvenient. It seems to me that one needs a good
>>>> reason (like a BIOS that understands the format and is able to boot
>>>> from it) to choose it.
>>>> 
>>> 
>>> Not really; it's actually a very simple table.
>>> 
>> 
>> It would be nice to know when this is going to make it in for my Linux 
>> projects.
>> I am running a 3Ware 9500 series with 3.1 TB disks.  I am able to use all 
>> the
>> storage at present with dsfs.  dsfs can support volumes up to 281 TB at 
>> present
>> but linux readir() can get into some problems when directories get really 
>> large.
>> 
>> I am not seeing problems with files that are 1.5 TB in size.  Have not 
>> tried to
>> create a 3TB file yet, but in theory, the VFS looks to support it.  I am 
>> getting around the
>> partition problem by basically ignoring the table extents (fdisk is broken 
>> with these large
>> partitions and wraps back to 700GB) if I have only created a single 
>> partition, I just query
>> the drive geometry and take the remaining space on the device and I ignore 
>> the partition
>> table.  It works fine.  If I detect more than one of my partitions I revert 
>> back to the actual
>> partition dimensions.
>> For Jens edification, I am using the BIO subsystem with this and I am 
>> seeing no problems
>> reading and writing these huge drives, so I think Linux 2.6.9 and 2.6.10 
>> will support this
>> well, and appears to.  I will be testing a combined striped array at around 
>> 20TB with multiple
>> controllers and FC/AL and will update if any problems are encountered.
>> Other than the partition problem, the base kernel seems to support these 
>> huge sizes with
>> 64 bit LBA addressing very well.
>> 
>> Jeff
>> 
>
> One other item I noticed is that the compiler for X86 has some problems doing 
> math for a 64 bit target
> variable, so when you are using Large LBA and doing something like:
>
> sector_t lba = part.start_lba + (block * (block_size /  sector_size));


I think the writer forgot about promotion rules so it's not
a compiler problem.

>
> you need to cast the variables if they are defined as 32 bit numbers because 
> the compiler is too stupid
> to realize you are adding the cumlative result into a 64-bit value, and it 
> will wrap the offset as a 32 bit number.
>

If block_size = uint32_t, and sector_size = uint32_t, then
block_size / sector_size is uint32_t, nothing more. Now, we have
a uint32_t * another uint32_t which is uint32_t with a possible
wrap, still correct. Then we have uint32_t + uint32_t which will
be promoted to sector_t (uint64_t).

If you need to promote variables ahead of the final assignment, the
writer needs to do this, not the compiler.


> i.e. 
> sector_t lba = part.start_lba + (sector_t)((sector_t)block * 
> ((sector_t)block_size /  (sector_t)sector_size));
>
> This works but if you leave off the type casting on any of the variables the 
> number reverts to a 32 bit value
> and wraps when you are calculating a 64 bit lba address. 
> Jeff
>

Cheers,
Dick Johnson
Penguin : Linux version 2.6.9 on an i686 machine (5537.79 BogoMips).
  Notice : All mail here is now cached for review by Dictator Bush.
                  98.36% of all statistics are fiction.