[lkml]   [2001]   [Nov]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: blocks or KB? (was: .. current meaning of blk_size array)
"Peter T. Breuer" wrote:
> "Martin Dalecki wrote:"
> > "Peter T. Breuer" wrote:
> > >
> > > Let me put it more plainly. Martin Daleki + rumour assures me that the
> > > blk_size array nowadays measure in blocks not KB, yet to me it seems that
> >
> > sectors = 512 per default
> > blocks = 1024 per default.
> I know that! But it's irrelevant. What I need to know is if blk_size is
> still counting in KB, or if it has switched to blocks.
> > Never said anything else.
> Err .. you said that blk_size is now measured in blocks, not KB. You
> said thet the rumour is true.
> "A month of sundays ago Martin Dalecki wrote:"
> > "Peter T. Breuer" wrote:
> > > Is blk_size[][] supposed to contain the size in KB or blocks?
> > There is no rumor it's in blocks.
> Maybe I misinterpret what you write. I interpret it as meaning "the
> rumour is not a rumour but a fact. It is in blocks".
> > Look at the initialization point for the arrays. They all use constants
> > which you can look up in the kernel headers.
> I _know_ that. It's irrelevant.
> The point is that if blk_size counts in KB, then the size of a device
> cannot reach more that 2^32 * 2^10 = 2^42 = 4TB. I'd personally say
> 2TB, becuase the int blk_size number is signed.
> That's rumoured not to be the case, and the max size of a device is
> supposed to be about 8 to 16TB. Let's suppose the rumour is true ..
> So we deduce that one has to assign a different meaning for the blk_size
> array. "count in blocks" is how the rumour goes. That way you can get
> 4 times higher sizes .. all the way to 8 or 16TB per device. And this
> is what is rumoured to be the case.
> Is it or is it not so? A straight answer from the list would be nice!
> > ./linux/fs.h:#define BLOCK_SIZE_BITS 10
> > ./linux/fs.h:#define BLOCK_SIZE (1<<BLOCK_SIZE_BITS)
> These are _defaults_ for _blksize_. Sure you can change it as you like,
> but according to the "blk_size is in KB" hypothesis, this matters not one
> iota to the size limit on devices. Change blksize and size does not
> change. But according to the "blk_size is in blocks" hypothesis, yes
> changing blksize will change the size of the device. Testing shows
> that scenario "blk_size is in KB" is true.
> Am I making plain the difference between blk_size and blksize?
> blk_size is the number of blocks or KB (which?) in a device. blksize is
> the size of the blocks. Is blk_size in KB or blocks?
> It should be in blocks if the size of a device is to reach 8 or 16TB.
> If it is in KB, we are limited to 2 or 4TB.
> > Which means 1024 bytes for blk_size as default value.
> But so what? That doesn't answer the question of whether blk_size
> is in blocks or not.
> > > it doesn't. Look at this code from ll_rw_blk.c in 2.4.[14]:
> Look at it:
> if (blk_size[major])
> minorsize = blk_size[major][MINOR(bh->b_rdev)];
> if (minorsize) {
> unsigned long maxsector = (minorsize << 1) + 1;
> This clearly hardcodes blk_size as measuring in units of 2 sectors, no
> matter what we set for blksize. It should be, in my view
> unsigned long maxsector =
> minorsize * blksize_size[major][MINOR(bh->b_rdev] + 1;
> or no device can be larger than 4TB. And neither can a filesystem, and
> neither can a file ...
> Now, I know I can write my own generic_make_request() code, but I have
> no intention of maintaining it through different kernel versions just
> to get the right size measurement. Besides, it's everyone's problem.
> Persuade me that this is not a bug, and an important one at that :-)
> Hellloooooo everybody! Linux cannot manage partitions greater than
> 4TB, ha ha ha hhhhhaaaa! ;-)
> I at least am getting up to devicesizes at the 8TB range.

The usage of it in block_dev.c is showing that in fact the matters
are more complicated that all your hypothesis together...

if (blk_size[MAJOR(dev)])
size = ((loff_t) blk_size[MAJOR(dev)][MINOR(dev)] << BLOCK_SIZE_BITS)
>> blocksize_bits;
size = INT_MAX;

The blk_size is in 90 out of 100 cases in units of 1024,
which is the default *logical* blocksize used by linux.
When this overflows, the block device layer just simply will not care
a damn bit about it and it will rely on the driver to notice overflow.
Therefore the answer is that yes it is in units of KB but linux will
still happy work with devices bigger then this.

Correct me please if I'm wrong... Slowly I start to look puzled myself.
If this is the case you can regard blk_size as the same kind of silly
blunt like the read_ahead array.

> Best wishes!
> Peter

- phone: +49 214 8656 283
- job: eVision-Ventures AG, LEV .de (MY OPINIONS ARE MY OWN!)
- langs: de_DE.ISO8859-1, en_US, pl_PL.ISO8859-2, last ressort:
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2005-03-22 13:13    [W:0.151 / U:0.184 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site