I would like to comment on the BTT docs a bit. There are some design points you might want to consider.
First, real use cases will have no read/write collisions. If you think of a file system, the case of reading a block that is being written or writing a single block twice just don't happen because the data itself is non deterministic. The driver still needs to handle these cases, but optimizing it for this is not all that logical.
Lets start at the BTT table. It would be useful if we could distinguish between a stable block and one that is getting updated now. An option is to encode the TRIM/ERROR bits as four "states" (stable, updating, trimmed, error) and just use the BTT entry as an index. This could probably point directly to the FLOG entry. The BTT table, at four bytes, has atomic updates without locks, so two threads can simultaneously update it to point to the FLOG table and then after the update, see if they won. If they did not win, they can wait for the first update to complete. The FLOG table could also have a parallel RAM based BTT2 table to store spinlocks or linked-lists to handle collisions. Then again, a simple spin or spin/sleep is probably good enough.
The same works for readers. If you read a block, check the BTT table after you finish the read. If it is the same, your read was good. If it changed underneath you, or is pointing to a FLOG block, then you need to wait or re-read. Again, the real-world frequency of collisions is very low. This would let you eliminate the RTT table entirely.
One final optimization would be to keep the BTT table both in standard RAM as well as in NV RAM. If standard RAM is faster, then reads could lookup blocks without touching the NV driver. For 512G, this is 1B blocks or 4G of RAM. Then again, if the NV RAM is just as fast, this would not help. Perhaps an option.
I have gotten into a lot of trouble optimizing for fio collisions when these collisions don't really impact real-workload performance. The code has to be "correct" in the collision case, but it does not really need to be fast.
Doug Dumitru
EasyCo LLC
On Fri, Jun 19, 2015 at 9:50 AM, Verma, Vishal L <vishal.l.verma@xxxxxxxxx> wrote:
Hi Mikulas,On Fri, 2015-06-19 at 12:33 -0400, Mikulas Patocka wrote:
> Hi
>
> I looked at the new the persistent memory block device driver
> (drivers/block/pmem.c and arch/x86/kernel/pmem.c) and it seems that the
> interface between them is incorrect.
>
> If I want to use persistent memory in another driver, for a different
> purpose, how can I make sure that that drivers/block/pmem.c doesn't attach
> to this piece of memory and export it? It seems not possible.
> drivers/block/pmem.c attaches to everything without regard that there may
> be other users of persistent memory.
>
> I think a correct solution would be to add a partition table at the
> beginning of persistent memory area and this partition table would
> describe which parts belong to which programs - so that different programs
> could use persistent memory and not step over each other's data. Is there
> some effort to standardize the partition table ongoing?
>
>
> BTW. some journaling filesystems assume that 512-byte sector is written
> atomically. drivers/block/pmem.c breaks this requirement. Persistent
> memory only gurantees 8-byte atomic writes.
I can answer this part - The idea is that file systems that need sector
atomicity will use the "Block Translation Table" (BTT) [1]. It would be
a stacked block device on top of a pmem device (or partition), and file
systems would be able to use it either for the entire space to get
atomicity for all blocks, or if they want to use DAX, make two
partitions, and enable the BTT only on one partition, and use it as the
logdev.
-Vishal
[1]: https://lkml.org/lkml/2015/6/17/950
>
> Mikulas
> _______________________________________________
> Linux-nvdimm mailing list
> Linux-nvdimm@xxxxxxxxxxxx
> https://lists.01.org/mailman/listinfo/linux-nvdimm
--
Doug Dumitru
EasyCo LLC
EasyCo LLC
-- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
