On Fri, 2014-01-31 at 16:20 -0500, Chris Mason wrote: > On 01/31/2014 02:02 PM, James Bottomley wrote: > > It has been reported: > > > > http://marc.info/?t=139111447200006 > > > > That large block devices (specifically devices > 16TB) crash when > > mounted on 32 bit systems. The problem specifically is that although > > CONFIG_LBDAF extends the size of sector_t within the block and storage > > layers to 64 bits, the buffer cache isn't big enough. Specifically, > > buffers are mapped through a single page cache mapping on the backing > > device inode. The size of the allowed offset in the page cache radix > > tree is pgoff_t which is 32 bits, so once the size of device goes beyond > > 16TB, this offset wraps and all hell breaks loose. > > > > The problem is that although the current single drive limit is about > > 4TB, it will only be a couple of years before 16TB devices are > > available. By then, I bet that most arm (and other exotic CPU) Linux > > based personal file servers are still going to be 32 bit, so they're not > > going to be able to take this generation (or beyond) of drives. The > > thing I'd like to discuss is how to fix this. There are several options > > I see, but there might be others. > > > > 1. Try to pretend that CONFIG_LBDAF is supposed to cap out at 16TB > > and there's nothing we can do about it ... this won't be at all > > popular with arm based file server manufacturers. > > 2. Slyly make sure that the buffer cache won't go over 16TB by > > keeping filesystem metadata below that limit ... the horse has > > probably already bolted on this one. > > 3. Increase pgoff_t and the radix tree indexes to u64 for > > CONFIG_LBDAF. This will blow out the size of struct page on 32 > > bits by 4 bytes and may have other knock on effects, but at > > least it will be transparent. > > 4. add an additional radix tree lookup within the buffer cache, so > > instead of a single inode for the buffer cache, we have a radix > > tree of them which are added and removed at the granularity of > > 16TB offsets as entries are requested. > > > > I started typing up that #3 is going to cause problems with RCU radix, > but it looks ok. I think we'll find a really scarey number of places > that interchange pgoff_t with unsigned long though. Yes, beyond the performance issues of doing 64 bits in the radix tree, it does look reasonably safe. > I prefer #4, but it means each FS needs to add code too. We assume > page_offset(page) maps to the disk in more than a few places. Hmm, yes, that's just a few cases of the readahead code, though, isn't it? The necessary fixes look fairly small per filesystem. James -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
