2011-10-12 18:42:22 +02:00
|
|
|
/*
|
|
|
|
* Copyright (C) 2011
|
|
|
|
* Boaz Harrosh <bharrosh@panasas.com>
|
|
|
|
*
|
|
|
|
* This file is part of the objects raid engine (ore).
|
|
|
|
*
|
|
|
|
* It is free software; you can redistribute it and/or modify
|
|
|
|
* it under the terms of the GNU General Public License version 2 as published
|
|
|
|
* by the Free Software Foundation.
|
|
|
|
*
|
|
|
|
* You should have received a copy of the GNU General Public License
|
|
|
|
* along with "ore". If not, write to the Free Software Foundation, Inc:
|
|
|
|
* "Free Software Foundation <info@fsf.org>"
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/gfp.h>
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
#include <linux/async_tx.h>
|
2011-10-12 18:42:22 +02:00
|
|
|
|
|
|
|
#include "ore_raid.h"
|
|
|
|
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
#undef ORE_DBGMSG2
|
|
|
|
#define ORE_DBGMSG2 ORE_DBGMSG
|
|
|
|
|
2014-02-09 14:03:15 +01:00
|
|
|
static struct page *_raid_page_alloc(void)
|
2011-10-12 18:42:22 +02:00
|
|
|
{
|
|
|
|
return alloc_page(GFP_KERNEL);
|
|
|
|
}
|
|
|
|
|
2014-02-09 14:03:15 +01:00
|
|
|
static void _raid_page_free(struct page *p)
|
2011-10-12 18:42:22 +02:00
|
|
|
{
|
|
|
|
__free_page(p);
|
|
|
|
}
|
|
|
|
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
/* This struct is forward declare in ore_io_state, but is private to here.
|
|
|
|
* It is put on ios->sp2d for RAID5/6 writes only. See _gen_xor_unit.
|
|
|
|
*
|
|
|
|
* __stripe_pages_2d is a 2d array of pages, and it is also a corner turn.
|
|
|
|
* Ascending page index access is sp2d(p-minor, c-major). But storage is
|
|
|
|
* sp2d[p-minor][c-major], so it can be properlly presented to the async-xor
|
|
|
|
* API.
|
|
|
|
*/
|
|
|
|
struct __stripe_pages_2d {
|
|
|
|
/* Cache some hot path repeated calculations */
|
|
|
|
unsigned parity;
|
|
|
|
unsigned data_devs;
|
|
|
|
unsigned pages_in_unit;
|
|
|
|
|
|
|
|
bool needed ;
|
|
|
|
|
|
|
|
/* Array size is pages_in_unit (layout->stripe_unit / PAGE_SIZE) */
|
|
|
|
struct __1_page_stripe {
|
|
|
|
bool alloc;
|
|
|
|
unsigned write_count;
|
|
|
|
struct async_submit_ctl submit;
|
|
|
|
struct dma_async_tx_descriptor *tx;
|
|
|
|
|
|
|
|
/* The size of this array is data_devs + parity */
|
|
|
|
struct page **pages;
|
|
|
|
struct page **scribble;
|
|
|
|
/* bool array, size of this array is data_devs */
|
|
|
|
char *page_is_read;
|
|
|
|
} _1p_stripes[];
|
|
|
|
};
|
|
|
|
|
|
|
|
/* This can get bigger then a page. So support multiple page allocations
|
|
|
|
* _sp2d_free should be called even if _sp2d_alloc fails (by returning
|
|
|
|
* none-zero).
|
|
|
|
*/
|
|
|
|
static int _sp2d_alloc(unsigned pages_in_unit, unsigned group_width,
|
|
|
|
unsigned parity, struct __stripe_pages_2d **psp2d)
|
|
|
|
{
|
|
|
|
struct __stripe_pages_2d *sp2d;
|
|
|
|
unsigned data_devs = group_width - parity;
|
|
|
|
struct _alloc_all_bytes {
|
|
|
|
struct __alloc_stripe_pages_2d {
|
|
|
|
struct __stripe_pages_2d sp2d;
|
|
|
|
struct __1_page_stripe _1p_stripes[pages_in_unit];
|
|
|
|
} __asp2d;
|
|
|
|
struct __alloc_1p_arrays {
|
|
|
|
struct page *pages[group_width];
|
|
|
|
struct page *scribble[group_width];
|
|
|
|
char page_is_read[data_devs];
|
|
|
|
} __a1pa[pages_in_unit];
|
|
|
|
} *_aab;
|
|
|
|
struct __alloc_1p_arrays *__a1pa;
|
|
|
|
struct __alloc_1p_arrays *__a1pa_end;
|
|
|
|
const unsigned sizeof__a1pa = sizeof(_aab->__a1pa[0]);
|
|
|
|
unsigned num_a1pa, alloc_size, i;
|
|
|
|
|
|
|
|
/* FIXME: check these numbers in ore_verify_layout */
|
|
|
|
BUG_ON(sizeof(_aab->__asp2d) > PAGE_SIZE);
|
|
|
|
BUG_ON(sizeof__a1pa > PAGE_SIZE);
|
|
|
|
|
|
|
|
if (sizeof(*_aab) > PAGE_SIZE) {
|
|
|
|
num_a1pa = (PAGE_SIZE - sizeof(_aab->__asp2d)) / sizeof__a1pa;
|
|
|
|
alloc_size = sizeof(_aab->__asp2d) + sizeof__a1pa * num_a1pa;
|
|
|
|
} else {
|
|
|
|
num_a1pa = pages_in_unit;
|
|
|
|
alloc_size = sizeof(*_aab);
|
|
|
|
}
|
|
|
|
|
|
|
|
_aab = kzalloc(alloc_size, GFP_KERNEL);
|
|
|
|
if (unlikely(!_aab)) {
|
|
|
|
ORE_DBGMSG("!! Failed to alloc sp2d size=%d\n", alloc_size);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
|
|
|
|
sp2d = &_aab->__asp2d.sp2d;
|
|
|
|
*psp2d = sp2d; /* From here Just call _sp2d_free */
|
|
|
|
|
|
|
|
__a1pa = _aab->__a1pa;
|
|
|
|
__a1pa_end = __a1pa + num_a1pa;
|
|
|
|
|
|
|
|
for (i = 0; i < pages_in_unit; ++i) {
|
|
|
|
if (unlikely(__a1pa >= __a1pa_end)) {
|
|
|
|
num_a1pa = min_t(unsigned, PAGE_SIZE / sizeof__a1pa,
|
|
|
|
pages_in_unit - i);
|
|
|
|
|
|
|
|
__a1pa = kzalloc(num_a1pa * sizeof__a1pa, GFP_KERNEL);
|
|
|
|
if (unlikely(!__a1pa)) {
|
|
|
|
ORE_DBGMSG("!! Failed to _alloc_1p_arrays=%d\n",
|
|
|
|
num_a1pa);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
__a1pa_end = __a1pa + num_a1pa;
|
|
|
|
/* First *pages is marked for kfree of the buffer */
|
|
|
|
sp2d->_1p_stripes[i].alloc = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
sp2d->_1p_stripes[i].pages = __a1pa->pages;
|
|
|
|
sp2d->_1p_stripes[i].scribble = __a1pa->scribble ;
|
|
|
|
sp2d->_1p_stripes[i].page_is_read = __a1pa->page_is_read;
|
|
|
|
++__a1pa;
|
|
|
|
}
|
|
|
|
|
|
|
|
sp2d->parity = parity;
|
|
|
|
sp2d->data_devs = data_devs;
|
|
|
|
sp2d->pages_in_unit = pages_in_unit;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void _sp2d_reset(struct __stripe_pages_2d *sp2d,
|
|
|
|
const struct _ore_r4w_op *r4w, void *priv)
|
|
|
|
{
|
|
|
|
unsigned data_devs = sp2d->data_devs;
|
|
|
|
unsigned group_width = data_devs + sp2d->parity;
|
2012-07-11 14:27:13 +02:00
|
|
|
int p, c;
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
|
|
|
|
if (!sp2d->needed)
|
|
|
|
return;
|
|
|
|
|
2012-07-11 14:27:13 +02:00
|
|
|
for (c = data_devs - 1; c >= 0; --c)
|
|
|
|
for (p = sp2d->pages_in_unit - 1; p >= 0; --p) {
|
|
|
|
struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
|
2012-07-11 14:27:13 +02:00
|
|
|
if (_1ps->page_is_read[c]) {
|
|
|
|
struct page *page = _1ps->pages[c];
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
|
2012-07-11 14:27:13 +02:00
|
|
|
r4w->put_page(priv, page);
|
|
|
|
_1ps->page_is_read[c] = false;
|
|
|
|
}
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
}
|
|
|
|
|
2012-07-11 14:27:13 +02:00
|
|
|
for (p = 0; p < sp2d->pages_in_unit; p++) {
|
|
|
|
struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
|
|
|
|
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
memset(_1ps->pages, 0, group_width * sizeof(*_1ps->pages));
|
|
|
|
_1ps->write_count = 0;
|
|
|
|
_1ps->tx = NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
sp2d->needed = false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void _sp2d_free(struct __stripe_pages_2d *sp2d)
|
|
|
|
{
|
|
|
|
unsigned i;
|
|
|
|
|
|
|
|
if (!sp2d)
|
|
|
|
return;
|
|
|
|
|
|
|
|
for (i = 0; i < sp2d->pages_in_unit; ++i) {
|
|
|
|
if (sp2d->_1p_stripes[i].alloc)
|
|
|
|
kfree(sp2d->_1p_stripes[i].pages);
|
|
|
|
}
|
|
|
|
|
|
|
|
kfree(sp2d);
|
|
|
|
}
|
|
|
|
|
|
|
|
static unsigned _sp2d_min_pg(struct __stripe_pages_2d *sp2d)
|
|
|
|
{
|
|
|
|
unsigned p;
|
|
|
|
|
|
|
|
for (p = 0; p < sp2d->pages_in_unit; p++) {
|
|
|
|
struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
|
|
|
|
|
|
|
|
if (_1ps->write_count)
|
|
|
|
return p;
|
|
|
|
}
|
|
|
|
|
|
|
|
return ~0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static unsigned _sp2d_max_pg(struct __stripe_pages_2d *sp2d)
|
|
|
|
{
|
2012-10-02 10:28:45 +02:00
|
|
|
int p;
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
|
|
|
|
for (p = sp2d->pages_in_unit - 1; p >= 0; --p) {
|
|
|
|
struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
|
|
|
|
|
|
|
|
if (_1ps->write_count)
|
|
|
|
return p;
|
|
|
|
}
|
|
|
|
|
|
|
|
return ~0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void _gen_xor_unit(struct __stripe_pages_2d *sp2d)
|
|
|
|
{
|
|
|
|
unsigned p;
|
|
|
|
for (p = 0; p < sp2d->pages_in_unit; p++) {
|
|
|
|
struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
|
|
|
|
|
|
|
|
if (!_1ps->write_count)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
init_async_submit(&_1ps->submit,
|
|
|
|
ASYNC_TX_XOR_ZERO_DST | ASYNC_TX_ACK,
|
2014-04-03 16:53:31 +02:00
|
|
|
NULL, NULL, NULL, (addr_conv_t *)_1ps->scribble);
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
|
|
|
|
/* TODO: raid6 */
|
|
|
|
_1ps->tx = async_xor(_1ps->pages[sp2d->data_devs], _1ps->pages,
|
|
|
|
0, sp2d->data_devs, PAGE_SIZE,
|
|
|
|
&_1ps->submit);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (p = 0; p < sp2d->pages_in_unit; p++) {
|
|
|
|
struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
|
|
|
|
/* NOTE: We wait for HW synchronously (I don't have such HW
|
|
|
|
* to test with.) Is parallelism needed with today's multi
|
|
|
|
* cores?
|
|
|
|
*/
|
|
|
|
async_tx_issue_pending(_1ps->tx);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void _ore_add_stripe_page(struct __stripe_pages_2d *sp2d,
|
|
|
|
struct ore_striping_info *si, struct page *page)
|
|
|
|
{
|
|
|
|
struct __1_page_stripe *_1ps;
|
|
|
|
|
|
|
|
sp2d->needed = true;
|
|
|
|
|
|
|
|
_1ps = &sp2d->_1p_stripes[si->cur_pg];
|
|
|
|
_1ps->pages[si->cur_comp] = page;
|
|
|
|
++_1ps->write_count;
|
|
|
|
|
|
|
|
si->cur_pg = (si->cur_pg + 1) % sp2d->pages_in_unit;
|
|
|
|
/* si->cur_comp is advanced outside at main loop */
|
|
|
|
}
|
|
|
|
|
2011-10-12 18:42:22 +02:00
|
|
|
void _ore_add_sg_seg(struct ore_per_dev_state *per_dev, unsigned cur_len,
|
|
|
|
bool not_last)
|
|
|
|
{
|
|
|
|
struct osd_sg_entry *sge;
|
|
|
|
|
|
|
|
ORE_DBGMSG("dev=%d cur_len=0x%x not_last=%d cur_sg=%d "
|
|
|
|
"offset=0x%llx length=0x%x last_sgs_total=0x%x\n",
|
|
|
|
per_dev->dev, cur_len, not_last, per_dev->cur_sg,
|
|
|
|
_LLU(per_dev->offset), per_dev->length,
|
|
|
|
per_dev->last_sgs_total);
|
|
|
|
|
|
|
|
if (!per_dev->cur_sg) {
|
|
|
|
sge = per_dev->sglist;
|
|
|
|
|
|
|
|
/* First time we prepare two entries */
|
|
|
|
if (per_dev->length) {
|
|
|
|
++per_dev->cur_sg;
|
|
|
|
sge->offset = per_dev->offset;
|
|
|
|
sge->len = per_dev->length;
|
|
|
|
} else {
|
|
|
|
/* Here the parity is the first unit of this object.
|
|
|
|
* This happens every time we reach a parity device on
|
|
|
|
* the same stripe as the per_dev->offset. We need to
|
|
|
|
* just skip this unit.
|
|
|
|
*/
|
|
|
|
per_dev->offset += cur_len;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
/* finalize the last one */
|
|
|
|
sge = &per_dev->sglist[per_dev->cur_sg - 1];
|
|
|
|
sge->len = per_dev->length - per_dev->last_sgs_total;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (not_last) {
|
|
|
|
/* Partly prepare the next one */
|
|
|
|
struct osd_sg_entry *next_sge = sge + 1;
|
|
|
|
|
|
|
|
++per_dev->cur_sg;
|
|
|
|
next_sge->offset = sge->offset + sge->len + cur_len;
|
|
|
|
/* Save cur len so we know how mutch was added next time */
|
|
|
|
per_dev->last_sgs_total = per_dev->length;
|
|
|
|
next_sge->len = 0;
|
|
|
|
} else if (!sge->len) {
|
|
|
|
/* Optimize for when the last unit is a parity */
|
|
|
|
--per_dev->cur_sg;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
static int _alloc_read_4_write(struct ore_io_state *ios)
|
|
|
|
{
|
|
|
|
struct ore_layout *layout = ios->layout;
|
|
|
|
int ret;
|
|
|
|
/* We want to only read those pages not in cache so worst case
|
|
|
|
* is a stripe populated with every other page
|
|
|
|
*/
|
|
|
|
unsigned sgs_per_dev = ios->sp2d->pages_in_unit + 2;
|
|
|
|
|
|
|
|
ret = _ore_get_io_state(layout, ios->oc,
|
|
|
|
layout->group_width * layout->mirrors_p1,
|
|
|
|
sgs_per_dev, 0, &ios->ios_read_4_write);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* @si contains info of the to-be-inserted page. Update of @si should be
|
|
|
|
* maintained by caller. Specificaly si->dev, si->obj_offset, ...
|
|
|
|
*/
|
2011-12-28 18:21:45 +01:00
|
|
|
static int _add_to_r4w(struct ore_io_state *ios, struct ore_striping_info *si,
|
|
|
|
struct page *page, unsigned pg_len)
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
{
|
|
|
|
struct request_queue *q;
|
|
|
|
struct ore_per_dev_state *per_dev;
|
|
|
|
struct ore_io_state *read_ios;
|
|
|
|
unsigned first_dev = si->dev - (si->dev %
|
|
|
|
(ios->layout->group_width * ios->layout->mirrors_p1));
|
|
|
|
unsigned comp = si->dev - first_dev;
|
|
|
|
unsigned added_len;
|
|
|
|
|
|
|
|
if (!ios->ios_read_4_write) {
|
|
|
|
int ret = _alloc_read_4_write(ios);
|
|
|
|
|
|
|
|
if (unlikely(ret))
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
read_ios = ios->ios_read_4_write;
|
|
|
|
read_ios->numdevs = ios->layout->group_width * ios->layout->mirrors_p1;
|
|
|
|
|
|
|
|
per_dev = &read_ios->per_dev[comp];
|
|
|
|
if (!per_dev->length) {
|
|
|
|
per_dev->bio = bio_kmalloc(GFP_KERNEL,
|
|
|
|
ios->sp2d->pages_in_unit);
|
|
|
|
if (unlikely(!per_dev->bio)) {
|
|
|
|
ORE_DBGMSG("Failed to allocate BIO size=%u\n",
|
|
|
|
ios->sp2d->pages_in_unit);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
per_dev->offset = si->obj_offset;
|
|
|
|
per_dev->dev = si->dev;
|
|
|
|
} else if (si->obj_offset != (per_dev->offset + per_dev->length)) {
|
|
|
|
u64 gap = si->obj_offset - (per_dev->offset + per_dev->length);
|
|
|
|
|
|
|
|
_ore_add_sg_seg(per_dev, gap, true);
|
|
|
|
}
|
|
|
|
q = osd_request_queue(ore_comp_dev(read_ios->oc, per_dev->dev));
|
2011-12-28 18:21:45 +01:00
|
|
|
added_len = bio_add_pc_page(q, per_dev->bio, page, pg_len,
|
|
|
|
si->obj_offset % PAGE_SIZE);
|
|
|
|
if (unlikely(added_len != pg_len)) {
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
ORE_DBGMSG("Failed to bio_add_pc_page bi_vcnt=%d\n",
|
|
|
|
per_dev->bio->bi_vcnt);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
|
2011-12-28 18:21:45 +01:00
|
|
|
per_dev->length += pg_len;
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2011-12-28 18:21:45 +01:00
|
|
|
/* read the beginning of an unaligned first page */
|
|
|
|
static int _add_to_r4w_first_page(struct ore_io_state *ios, struct page *page)
|
|
|
|
{
|
|
|
|
struct ore_striping_info si;
|
|
|
|
unsigned pg_len;
|
|
|
|
|
|
|
|
ore_calc_stripe_info(ios->layout, ios->offset, 0, &si);
|
|
|
|
|
|
|
|
pg_len = si.obj_offset % PAGE_SIZE;
|
|
|
|
si.obj_offset -= pg_len;
|
|
|
|
|
|
|
|
ORE_DBGMSG("offset=0x%llx len=0x%x index=0x%lx dev=%x\n",
|
|
|
|
_LLU(si.obj_offset), pg_len, page->index, si.dev);
|
|
|
|
|
|
|
|
return _add_to_r4w(ios, &si, page, pg_len);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* read the end of an incomplete last page */
|
|
|
|
static int _add_to_r4w_last_page(struct ore_io_state *ios, u64 *offset)
|
|
|
|
{
|
|
|
|
struct ore_striping_info si;
|
|
|
|
struct page *page;
|
|
|
|
unsigned pg_len, p, c;
|
|
|
|
|
|
|
|
ore_calc_stripe_info(ios->layout, *offset, 0, &si);
|
|
|
|
|
|
|
|
p = si.unit_off / PAGE_SIZE;
|
|
|
|
c = _dev_order(ios->layout->group_width * ios->layout->mirrors_p1,
|
|
|
|
ios->layout->mirrors_p1, si.par_dev, si.dev);
|
|
|
|
page = ios->sp2d->_1p_stripes[p].pages[c];
|
|
|
|
|
|
|
|
pg_len = PAGE_SIZE - (si.unit_off % PAGE_SIZE);
|
|
|
|
*offset += pg_len;
|
|
|
|
|
|
|
|
ORE_DBGMSG("p=%d, c=%d next-offset=0x%llx len=0x%x dev=%x par_dev=%d\n",
|
|
|
|
p, c, _LLU(*offset), pg_len, si.dev, si.par_dev);
|
|
|
|
|
|
|
|
BUG_ON(!page);
|
|
|
|
|
|
|
|
return _add_to_r4w(ios, &si, page, pg_len);
|
|
|
|
}
|
|
|
|
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
static void _mark_read4write_pages_uptodate(struct ore_io_state *ios, int ret)
|
|
|
|
{
|
|
|
|
struct bio_vec *bv;
|
|
|
|
unsigned i, d;
|
|
|
|
|
|
|
|
/* loop on all devices all pages */
|
|
|
|
for (d = 0; d < ios->numdevs; d++) {
|
|
|
|
struct bio *bio = ios->per_dev[d].bio;
|
|
|
|
|
|
|
|
if (!bio)
|
|
|
|
continue;
|
|
|
|
|
2013-02-06 21:23:11 +01:00
|
|
|
bio_for_each_segment_all(bv, bio, i) {
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
struct page *page = bv->bv_page;
|
|
|
|
|
|
|
|
SetPageUptodate(page);
|
|
|
|
if (PageError(page))
|
|
|
|
ClearPageError(page);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* read_4_write is hacked to read the start of the first stripe and/or
|
|
|
|
* the end of the last stripe. If needed, with an sg-gap at each device/page.
|
|
|
|
* It is assumed to be called after the to_be_written pages of the first stripe
|
|
|
|
* are populating ios->sp2d[][]
|
|
|
|
*
|
|
|
|
* NOTE: We call ios->r4w->lock_fn for all pages needed for parity calculations
|
|
|
|
* These pages are held at sp2d[p].pages[c] but with
|
|
|
|
* sp2d[p].page_is_read[c] = true. At _sp2d_reset these pages are
|
|
|
|
* ios->r4w->lock_fn(). The ios->r4w->lock_fn might signal that the page is
|
|
|
|
* @uptodate=true, so we don't need to read it, only unlock, after IO.
|
|
|
|
*
|
|
|
|
* TODO: The read_4_write should calc a need_to_read_pages_count, if bigger then
|
|
|
|
* to-be-written count, we should consider the xor-in-place mode.
|
|
|
|
* need_to_read_pages_count is the actual number of pages not present in cache.
|
|
|
|
* maybe "devs_in_group - ios->sp2d[p].write_count" is a good enough
|
|
|
|
* approximation? In this mode the read pages are put in the empty places of
|
|
|
|
* ios->sp2d[p][*], xor is calculated the same way. These pages are
|
|
|
|
* allocated/freed and don't go through cache
|
|
|
|
*/
|
ore: Fix NFS crash by supporting any unaligned RAID IO
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>
2012-06-08 00:19:07 +02:00
|
|
|
static int _read_4_write_first_stripe(struct ore_io_state *ios)
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
{
|
|
|
|
struct ore_striping_info read_si;
|
|
|
|
struct __stripe_pages_2d *sp2d = ios->sp2d;
|
|
|
|
u64 offset = ios->si.first_stripe_start;
|
ore: Fix NFS crash by supporting any unaligned RAID IO
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>
2012-06-08 00:19:07 +02:00
|
|
|
unsigned c, p, min_p = sp2d->pages_in_unit, max_p = -1;
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
|
|
|
|
if (offset == ios->offset) /* Go to start collect $200 */
|
|
|
|
goto read_last_stripe;
|
|
|
|
|
|
|
|
min_p = _sp2d_min_pg(sp2d);
|
|
|
|
max_p = _sp2d_max_pg(sp2d);
|
|
|
|
|
ore: Fix NFS crash by supporting any unaligned RAID IO
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>
2012-06-08 00:19:07 +02:00
|
|
|
ORE_DBGMSG("stripe_start=0x%llx ios->offset=0x%llx min_p=%d max_p=%d\n",
|
|
|
|
offset, ios->offset, min_p, max_p);
|
|
|
|
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
for (c = 0; ; c++) {
|
|
|
|
ore_calc_stripe_info(ios->layout, offset, 0, &read_si);
|
|
|
|
read_si.obj_offset += min_p * PAGE_SIZE;
|
|
|
|
offset += min_p * PAGE_SIZE;
|
|
|
|
for (p = min_p; p <= max_p; p++) {
|
|
|
|
struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
|
|
|
|
struct page **pp = &_1ps->pages[c];
|
|
|
|
bool uptodate;
|
|
|
|
|
2011-12-28 18:21:45 +01:00
|
|
|
if (*pp) {
|
|
|
|
if (ios->offset % PAGE_SIZE)
|
|
|
|
/* Read the remainder of the page */
|
|
|
|
_add_to_r4w_first_page(ios, *pp);
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
/* to-be-written pages start here */
|
|
|
|
goto read_last_stripe;
|
2011-12-28 18:21:45 +01:00
|
|
|
}
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
|
|
|
|
*pp = ios->r4w->get_page(ios->private, offset,
|
|
|
|
&uptodate);
|
|
|
|
if (unlikely(!*pp))
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
if (!uptodate)
|
2011-12-28 18:21:45 +01:00
|
|
|
_add_to_r4w(ios, &read_si, *pp, PAGE_SIZE);
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
|
|
|
|
/* Mark read-pages to be cache_released */
|
|
|
|
_1ps->page_is_read[c] = true;
|
|
|
|
read_si.obj_offset += PAGE_SIZE;
|
|
|
|
offset += PAGE_SIZE;
|
|
|
|
}
|
|
|
|
offset += (sp2d->pages_in_unit - p) * PAGE_SIZE;
|
|
|
|
}
|
|
|
|
|
|
|
|
read_last_stripe:
|
ore: Fix NFS crash by supporting any unaligned RAID IO
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>
2012-06-08 00:19:07 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int _read_4_write_last_stripe(struct ore_io_state *ios)
|
|
|
|
{
|
|
|
|
struct ore_striping_info read_si;
|
|
|
|
struct __stripe_pages_2d *sp2d = ios->sp2d;
|
|
|
|
u64 offset;
|
|
|
|
u64 last_stripe_end;
|
|
|
|
unsigned bytes_in_stripe = ios->si.bytes_in_stripe;
|
|
|
|
unsigned c, p, min_p = sp2d->pages_in_unit, max_p = -1;
|
|
|
|
|
2011-12-28 18:21:45 +01:00
|
|
|
offset = ios->offset + ios->length;
|
|
|
|
if (offset % PAGE_SIZE)
|
|
|
|
_add_to_r4w_last_page(ios, &offset);
|
|
|
|
/* offset will be aligned to next page */
|
|
|
|
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
last_stripe_end = div_u64(offset + bytes_in_stripe - 1, bytes_in_stripe)
|
|
|
|
* bytes_in_stripe;
|
|
|
|
if (offset == last_stripe_end) /* Optimize for the aligned case */
|
|
|
|
goto read_it;
|
|
|
|
|
|
|
|
ore_calc_stripe_info(ios->layout, offset, 0, &read_si);
|
|
|
|
p = read_si.unit_off / PAGE_SIZE;
|
|
|
|
c = _dev_order(ios->layout->group_width * ios->layout->mirrors_p1,
|
|
|
|
ios->layout->mirrors_p1, read_si.par_dev, read_si.dev);
|
|
|
|
|
|
|
|
if (min_p == sp2d->pages_in_unit) {
|
|
|
|
/* Didn't do it yet */
|
|
|
|
min_p = _sp2d_min_pg(sp2d);
|
|
|
|
max_p = _sp2d_max_pg(sp2d);
|
|
|
|
}
|
|
|
|
|
ore: Fix NFS crash by supporting any unaligned RAID IO
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>
2012-06-08 00:19:07 +02:00
|
|
|
ORE_DBGMSG("offset=0x%llx stripe_end=0x%llx min_p=%d max_p=%d\n",
|
|
|
|
offset, last_stripe_end, min_p, max_p);
|
|
|
|
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
while (offset < last_stripe_end) {
|
|
|
|
struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
|
|
|
|
|
|
|
|
if ((min_p <= p) && (p <= max_p)) {
|
|
|
|
struct page *page;
|
|
|
|
bool uptodate;
|
|
|
|
|
|
|
|
BUG_ON(_1ps->pages[c]);
|
|
|
|
page = ios->r4w->get_page(ios->private, offset,
|
|
|
|
&uptodate);
|
|
|
|
if (unlikely(!page))
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
_1ps->pages[c] = page;
|
|
|
|
/* Mark read-pages to be cache_released */
|
|
|
|
_1ps->page_is_read[c] = true;
|
|
|
|
if (!uptodate)
|
2011-12-28 18:21:45 +01:00
|
|
|
_add_to_r4w(ios, &read_si, page, PAGE_SIZE);
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
offset += PAGE_SIZE;
|
|
|
|
if (p == (sp2d->pages_in_unit - 1)) {
|
|
|
|
++c;
|
|
|
|
p = 0;
|
|
|
|
ore_calc_stripe_info(ios->layout, offset, 0, &read_si);
|
|
|
|
} else {
|
|
|
|
read_si.obj_offset += PAGE_SIZE;
|
|
|
|
++p;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
read_it:
|
ore: Fix NFS crash by supporting any unaligned RAID IO
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>
2012-06-08 00:19:07 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int _read_4_write_execute(struct ore_io_state *ios)
|
|
|
|
{
|
|
|
|
struct ore_io_state *ios_read;
|
|
|
|
unsigned i;
|
|
|
|
int ret;
|
|
|
|
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
ios_read = ios->ios_read_4_write;
|
|
|
|
if (!ios_read)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
/* FIXME: Ugly to signal _sbi_read_mirror that we have bio(s). Change
|
|
|
|
* to check for per_dev->bio
|
|
|
|
*/
|
|
|
|
ios_read->pages = ios->pages;
|
|
|
|
|
|
|
|
/* Now read these devices */
|
|
|
|
for (i = 0; i < ios_read->numdevs; i += ios_read->layout->mirrors_p1) {
|
|
|
|
ret = _ore_read_mirror(ios_read, i);
|
|
|
|
if (unlikely(ret))
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
ret = ore_io_execute(ios_read); /* Synchronus execution */
|
|
|
|
if (unlikely(ret)) {
|
|
|
|
ORE_DBGMSG("!! ore_io_execute => %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
_mark_read4write_pages_uptodate(ios_read, ret);
|
ore: Fix NFS crash by supporting any unaligned RAID IO
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>
2012-06-08 00:19:07 +02:00
|
|
|
ore_put_io_state(ios_read);
|
|
|
|
ios->ios_read_4_write = NULL; /* Might need a reuse at last stripe */
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2011-10-12 18:42:22 +02:00
|
|
|
/* In writes @cur_len means length left. .i.e cur_len==0 is the last parity U */
|
|
|
|
int _ore_add_parity_unit(struct ore_io_state *ios,
|
|
|
|
struct ore_striping_info *si,
|
|
|
|
struct ore_per_dev_state *per_dev,
|
|
|
|
unsigned cur_len)
|
|
|
|
{
|
|
|
|
if (ios->reading) {
|
2011-12-28 18:14:23 +01:00
|
|
|
if (per_dev->cur_sg >= ios->sgs_per_dev) {
|
|
|
|
ORE_DBGMSG("cur_sg(%d) >= sgs_per_dev(%d)\n" ,
|
|
|
|
per_dev->cur_sg, ios->sgs_per_dev);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
2011-10-12 18:42:22 +02:00
|
|
|
_ore_add_sg_seg(per_dev, cur_len, true);
|
|
|
|
} else {
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
struct __stripe_pages_2d *sp2d = ios->sp2d;
|
2011-10-12 18:42:22 +02:00
|
|
|
struct page **pages = ios->parity_pages + ios->cur_par_page;
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
unsigned num_pages;
|
2011-10-12 18:42:22 +02:00
|
|
|
unsigned array_start = 0;
|
|
|
|
unsigned i;
|
|
|
|
int ret;
|
|
|
|
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
si->cur_pg = _sp2d_min_pg(sp2d);
|
|
|
|
num_pages = _sp2d_max_pg(sp2d) + 1 - si->cur_pg;
|
|
|
|
|
|
|
|
if (!cur_len) /* If last stripe operate on parity comp */
|
|
|
|
si->cur_comp = sp2d->data_devs;
|
|
|
|
|
|
|
|
if (!per_dev->length) {
|
|
|
|
per_dev->offset += si->cur_pg * PAGE_SIZE;
|
|
|
|
/* If first stripe, Read in all read4write pages
|
|
|
|
* (if needed) before we calculate the first parity.
|
|
|
|
*/
|
ore: Fix NFS crash by supporting any unaligned RAID IO
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>
2012-06-08 00:19:07 +02:00
|
|
|
_read_4_write_first_stripe(ios);
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
}
|
ore: Fix NFS crash by supporting any unaligned RAID IO
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>
2012-06-08 00:19:07 +02:00
|
|
|
if (!cur_len) /* If last stripe r4w pages of last stripe */
|
|
|
|
_read_4_write_last_stripe(ios);
|
|
|
|
_read_4_write_execute(ios);
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
|
2011-10-12 18:42:22 +02:00
|
|
|
for (i = 0; i < num_pages; i++) {
|
|
|
|
pages[i] = _raid_page_alloc();
|
|
|
|
if (unlikely(!pages[i]))
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
++(ios->cur_par_page);
|
|
|
|
}
|
|
|
|
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
BUG_ON(si->cur_comp != sp2d->data_devs);
|
|
|
|
BUG_ON(si->cur_pg + num_pages > sp2d->pages_in_unit);
|
2011-10-12 18:42:22 +02:00
|
|
|
|
|
|
|
ret = _ore_add_stripe_unit(ios, &array_start, 0, pages,
|
|
|
|
per_dev, num_pages * PAGE_SIZE);
|
|
|
|
if (unlikely(ret))
|
|
|
|
return ret;
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
|
|
|
|
/* TODO: raid6 if (last_parity_dev) */
|
|
|
|
_gen_xor_unit(sp2d);
|
|
|
|
_sp2d_reset(sp2d, ios->r4w, ios->private);
|
2011-10-12 18:42:22 +02:00
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
int _ore_post_alloc_raid_stuff(struct ore_io_state *ios)
|
|
|
|
{
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
if (ios->parity_pages) {
|
ore: Fix NFS crash by supporting any unaligned RAID IO
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>
2012-06-08 00:19:07 +02:00
|
|
|
struct ore_layout *layout = ios->layout;
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
unsigned pages_in_unit = layout->stripe_unit / PAGE_SIZE;
|
|
|
|
|
|
|
|
if (_sp2d_alloc(pages_in_unit, layout->group_width,
|
|
|
|
layout->parity, &ios->sp2d)) {
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
}
|
2011-10-12 18:42:22 +02:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
void _ore_free_raid_stuff(struct ore_io_state *ios)
|
|
|
|
{
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
if (ios->sp2d) { /* writing and raid */
|
2011-10-12 18:42:22 +02:00
|
|
|
unsigned i;
|
|
|
|
|
|
|
|
for (i = 0; i < ios->cur_par_page; i++) {
|
|
|
|
struct page *page = ios->parity_pages[i];
|
|
|
|
|
|
|
|
if (page)
|
|
|
|
_raid_page_free(page);
|
|
|
|
}
|
|
|
|
if (ios->extra_part_alloc)
|
|
|
|
kfree(ios->parity_pages);
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
/* If IO returned an error pages might need unlocking */
|
|
|
|
_sp2d_reset(ios->sp2d, ios->r4w, ios->private);
|
|
|
|
_sp2d_free(ios->sp2d);
|
2011-10-12 18:42:22 +02:00
|
|
|
} else {
|
|
|
|
/* Will only be set if raid reading && sglist is big */
|
|
|
|
if (ios->extra_part_alloc)
|
|
|
|
kfree(ios->per_dev[0].sglist);
|
|
|
|
}
|
ore: RAID5 Write
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>
2011-10-14 15:33:51 +02:00
|
|
|
if (ios->ios_read_4_write)
|
|
|
|
ore_put_io_state(ios->ios_read_4_write);
|
2011-10-12 18:42:22 +02:00
|
|
|
}
|