The read-4-write pages are locked in address ascending order.
But where unlocked in a way easiest for coding. Fix that,
locks should be released in opposite order of locking, .i.e
descending address order.
I have not hit this dead-lock. It was found by inspecting the
dbug print-outs. I suspect there is an higher lock at caller that
protects us, but fix it regardless.
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>
In RAID_5/6 We used to not permit an IO that it's end
byte is not stripe_size aligned and spans more than one stripe.
.i.e the caller must check if after submission the actual
transferred bytes is shorter, and would need to resubmit
a new IO with the remainder.
Exofs supports this, and NFS was supposed to support this
as well with it's short write mechanism. But late testing has
exposed a CRASH when this is used with none-RPC layout-drivers.
The change at NFS is deep and risky, in it's place the fix
at ORE to lift the limitation is actually clean and simple.
So here it is below.
The principal here is that in the case of unaligned IO on
both ends, beginning and end, we will send two read requests
one like old code, before the calculation of the first stripe,
and also a new site, before the calculation of the last stripe.
If any "boundary" is aligned or the complete IO is within a single
stripe. we do a single read like before.
The code is clean and simple by splitting the old _read_4_write
into 3 even parts:
1._read_4_write_first_stripe
2. _read_4_write_last_stripe
3. _read_4_write_execute
And calling 1+3 at the same place as before. 2+3 before last
stripe, and in the case of all in a single stripe then 1+2+3
is preformed additively.
Why did I not think of it before. Well I had a strike of
genius because I have stared at this code for 2 years, and did
not find this simple solution, til today. Not that I did not try.
This solution is much better for NFS than the previous supposedly
solution because the short write was dealt with out-of-band after
IO_done, which would cause for a seeky IO pattern where as in here
we execute in order. At both solutions we do 2 separate reads, only
here we do it within a single IO request. (And actually combine two
writes into a single submission)
NFS/exofs code need not change since the ORE API communicates the new
shorter length on return, what will happen is that this case would not
occur anymore.
hurray!!
[Stable this is an NFS bug since 3.2 Kernel should apply cleanly]
CC: Stable Tree <stable@kernel.org>
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>
NFS might send us offsets that are not PAGE aligned. So
we must read in the reminder of the first/last pages, in cases
we need it for Parity calculations.
We only add an sg segments to read the partial page. But
we don't mark it as read=true because it is a lock-for-write
page.
TODO: In some cases (IO spans a single unit) we can just
adjust the raid_unit offset/length, but this is left for
later Kernels.
[Bug in 3.2.0 Kernel]
CC: Stable Tree <stable@kernel.org>
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>
When reading RAID5 files, in rare cases, we calculated too
few sg segments. There should be two extra for the beginning
and end partial units.
Also "too few sg segments" should not be a BUG_ON there is
all the mechanics in place to handle it, as a short read.
So just return -ENOMEM and the rest of the code will gracefully
split the IO.
[Bug in 3.2.0 Kernel]
CC: Stable Tree <stable@kernel.org>
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>
This is finally the RAID5 Write support.
The bigger part of this patch is not the XOR engine itself, But the
read4write logic, which is a complete mini prepare_for_striping
reading engine that can read scattered pages of a stripe into cache
so it can be used for XOR calculation. That is, if the write was not
stripe aligned.
The main algorithm behind the XOR engine is the 2 dimensional array:
struct __stripe_pages_2d.
A drawing might save 1000 words
---
__stripe_pages_2d
|
n = pages_in_stripe_unit;
w = group_width - parity;
| pages array presented to the XOR lib
| |
V |
__1_page_stripe[0].pages --> [c0][c1]..[cw][c_par] <---|
| |
__1_page_stripe[1].pages --> [c0][c1]..[cw][c_par] <---
|
... | ...
|
__1_page_stripe[n].pages --> [c0][c1]..[cw][c_par]
^
|
data added columns first then row
---
The pages are put on this array columns first. .i.e:
p0-of-c0, p1-of-c0, ... pn-of-c0, p0-of-c1, ...
So we are doing a corner turn of the pages.
Note that pages will zigzag down and left. but are put sequentially
in growing order. So when the time comes to XOR the stripe, only the
beginning and end of the array need be checked. We scan the array
and any NULL spot will be field by pages-to-be-read.
The FS that wants to support RAID5 needs to supply an
operations-vector that searches a given page in cache, and specifies
if the page is uptodate or need reading. All these pages to be read
are put on a slave ore_io_state and synchronously read. All the pages
of a stripe are read in one IO, using the scatter gather mechanism.
In write we constrain our IO to only be incomplete on a single
stripe. Meaning either the complete IO is within a single stripe so
we might have pages to read from both beginning or end of the
strip. Or we have some reading to do at beginning but end at strip
boundary. The left over pages are pushed to the next IO by the API
already established by previous work, where an IO offset/length
combination presented to the ORE might get the length truncated and
the user must re-submit the leftover pages. (Both exofs and NFS
support this)
But any ORE user should make it's best effort to align it's IO
before hand and avoid complications. A cached ore_layout->stripe_size
member can be used for that calculation. (NOTE: that ORE demands
that stripe_size may not be bigger then 32bit)
What else? Well read it and tell me.
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>
This patch introduces the first stage of RAID5 support
mainly the skip-over-raid-units when reading. For
writes it inserts BLANK units, into where XOR blocks
should be calculated and written to.
It introduces the new "general raid maths", and the main
additional parameters and components needed for raid5.
Since at this stage it could corrupt future version that
actually do support raid5. The enablement of raid5
mounting and setting of parity-count > 0 is disabled. So
the raid5 code will never be used. Mounting of raid5 is
only enabled later once the basic XOR write is also in.
But if the patch "enable RAID5" is applied this code has
been tested to be able to properly read raid5 volumes
and is according to standard.
Also it has been tested that the new maths still properly
supports RAID0 and grouping code just as before.
(BTW: I have found more bugs in the pnfs-obj RAID math
fixed here)
The ore.c file is getting too big, so new ore_raid.[hc]
files are added that will include the special raid stuff
that are not used in striping and mirrors. In future write
support these will get bigger.
When adding the ore_raid.c to Kbuild file I was forced to
rename ore.ko to libore.ko. Is it possible to keep source
file, say ore.c and module file ore.ko the same even if there
are multiple files inside ore.ko?
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>
