I am used to doing all of this strictly in DRAM, so my optimization
background has differing overhead points.
On Mon, Jun 22, 2015 at 2:25 PM, Verma, Vishal L <vishal.l.verma(a)intel.com>
On Mon, 2015-06-22 at 09:50 -0700, Doug Dumitru wrote:
> I would like to comment on the BTT docs a bit. There are some design
> points you might want to consider.
Thanks for the design review -
> 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.
Agreed - we did try to keep this in mind, and I still have a list of
experiments to try that might further reduce the impact of collision
handling on normal IO.
> 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.
So first off, this made me realize the documentation was slightly dated
w.r.t. the use of the 'E' and 'Z' flags - we currently (and the code
reflects this) do use all four states from those two bits. (The next
posting of this patchset should reflect this change).
31 - 30 : Error and Zero flags - Used in the following way:
00 Initial state. Reads return zeroes; Premap = Postmap
01 Zero state: Reads return zeroes
10 Error state: Reads fail; Writes clear 'E' bit
11 Normal Block – has valid postmap
29 - 0 : Mappings to internal 'postmap' blocks
Apart from this, I didn't quite understand what you meant by:
> ... and just use the BTT entry as an index. This could probably point
directly to the FLOG entry.
Care to elaborate with an example maybe?
I am not sure how "committed" to binary region sizes here, but the first
thing that sticks out is that all but the normal state don't really need an
index entry. If you are willing to suffer some "non binary divide"
operations, you can increase the efficiency of your region by using these 4
0 - Initial state
1 - Zero/trim/discard block, but why not just '0' above
2 - Error
3-F - Future flags
10-FFFF - index into FLOG table
10000-FFFFFFFF - index directly into RAM store
This pushes the region to just under 4B entries (2TB with 512 and 16TB
with 4K). If you must stay binary, consider a single bit.
0 - Initial state
1 - Zero
2 - Error
10-FFFF - index into FLOW table
8FFFFFFFF - FFFFFFFF - index directly into RAM store
The FLOG table can have pre/post RAM store index (per your current
design), so a read on a write has to do a double lookup, but that might be
good anyway, and the double lookup is only on a collision anyway.
In terms of read overlaps, you could argue you can do this without a lock.
If you read a BTT entry and it is 1000-FFFFFFFF, and you deterministically
control how often a block gets re-used, it is safe to just read even if an
update starts half-way through. If you want to be paranoid, setup an
atomic for updates that the read can use to protect itself.
read update ATOMIC
if BTT is not an index into the FLOG table then read data
read update ATOMIC again
If ATOMIC has not moved more than the guaranteed blocks over commit number,
you know the read is valid data and the block has not been overwritten.
The write logic just increments the ATOMIC (and lets it wrap around)
whenever a new block is used. As long as the avail block first-in
first-out list is slightly bigger than the advertised device capacity, you
can guarantee "N" updates before a read gets the rug pulled out from under
> 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
There are a couple of issues I see with this -
1. This could, theoretically, suffer from the A-B-A problem - a writer
could update the same mapping twice, ending up with an apparently
untouched map entry, but the data has gone through churn while the
reader was still reading it.
2. Writing/reading the RTT is cheap as it is entirely in DRAM memory,
whereas another check on the map would mean NVDIMM IO, which could
potentially be slow enough to offset any benefit from getting rid of the
If the API read is expensive, then you are probably right. Consider an
ATOMIC DRAM counter as described above. This requires you to keep a small
number of blocks "spare" so that reads complete before they can get
clobbered, but this lets reads run lock free.
One of the pending tests I plan to do is to make the RTT a hash map
instead of a list. i.e. Reader adds a refcount to rtt[postmap ABA], and
writer checks just RTT[free-block]. This keeps things the same for the
reader, but for writers, they are sapred from walking a list, at the
cost of false hash collisions. Obviously, with this path we'd have to
make the RTT a lot bigger than it currently is - and we'd need to figure
out if the increased memory footprint, and hash collisions outweigh the
cost of walking an in-memory list..
If the BTT is a pointer to the FLOG, I think the RTT can go away 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.
We've thought about this, and while it could help, the added complexity
of keeping the two in sync may make it debatable - we'll have to
re-evaluate once NVDIMMs become more easily available :)
It should just be a double update of the same four bytes to both places.
Update the RAM last but only ACK the write after both are written and
memory fenced. If a read comes along and gets the RAM image out of sync
with the NV image, you are reading "during" a write so getting the previous
data is still "correct"
as you started the read before the write ACKd.
> 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(a)intel.com> wrote:
> On Fri, 2015-06-19 at 12:33 -0400, Mikulas Patocka wrote:
> > Hi
> > I looked at the new the persistent memory block device
> > (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
> > 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
> > 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
> > 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.
> > memory only gurantees 8-byte atomic writes.
> Hi Mikulas,
> I can answer this part - The idea is that file systems that
> need sector
> atomicity will use the "Block Translation Table" (BTT) . 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
> 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
> : https://lkml.org/lkml/2015/6/17/950
> > Mikulas
> > _______________________________________________
> > Linux-nvdimm mailing list
> > Linux-nvdimm(a)lists.01.org
> > https://lists.01.org/mailman/listinfo/linux-nvdimm
> dm-devel mailing list
> Doug Dumitru
> EasyCo LLC
> Linux-nvdimm mailing list
My comment on block sizes was only 512/4K, but the code should be easy to
setup at anything binary.