2005-04-16 22:20:36 +00:00
/*
* JFFS2 - - Journalling Flash File System , Version 2.
*
* Copyright ( C ) 2001 - 2003 Red Hat , Inc .
*
* Created by David Woodhouse < dwmw2 @ infradead . org >
*
* For licensing information , see the file ' LICENCE ' in this directory .
*
2005-11-07 11:16:07 +00:00
* $ Id : readinode . c , v 1.143 2005 / 11 / 07 11 : 14 : 41 gleixner Exp $
2005-04-16 22:20:36 +00:00
*
*/
# include <linux/kernel.h>
2005-07-30 15:29:30 +00:00
# include <linux/sched.h>
2005-04-16 22:20:36 +00:00
# include <linux/slab.h>
# include <linux/fs.h>
# include <linux/crc32.h>
# include <linux/pagemap.h>
# include <linux/mtd/mtd.h>
# include <linux/compiler.h>
# include "nodelist.h"
2005-08-01 12:05:22 +00:00
/*
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
* Check the data CRC of the node .
*
* Returns : 0 if the data CRC is correct ;
* 1 - if incorrect ;
* error code if an error occured .
*/
static int check_node_data ( struct jffs2_sb_info * c , struct jffs2_tmp_dnode_info * tn )
{
struct jffs2_raw_node_ref * ref = tn - > fn - > raw ;
int err = 0 , pointed = 0 ;
struct jffs2_eraseblock * jeb ;
unsigned char * buffer ;
uint32_t crc , ofs , len ;
size_t retlen ;
BUG_ON ( tn - > csize = = 0 ) ;
if ( ! jffs2_is_writebuffered ( c ) )
goto adj_acc ;
/* Calculate how many bytes were already checked */
ofs = ref_offset ( ref ) + sizeof ( struct jffs2_raw_inode ) ;
len = ofs % c - > wbuf_pagesize ;
if ( likely ( len ) )
len = c - > wbuf_pagesize - len ;
if ( len > = tn - > csize ) {
dbg_readinode ( " no need to check node at %#08x, data length %u, data starts at %#08x - it has already been checked. \n " ,
ref_offset ( ref ) , tn - > csize , ofs ) ;
goto adj_acc ;
}
ofs + = len ;
len = tn - > csize - len ;
dbg_readinode ( " check node at %#08x, data length %u, partial CRC %#08x, correct CRC %#08x, data starts at %#08x, start checking from %#08x - %u bytes. \n " ,
ref_offset ( ref ) , tn - > csize , tn - > partial_crc , tn - > data_crc , ofs - len , ofs , len ) ;
# ifndef __ECOS
/* TODO: instead, incapsulate point() stuff to jffs2_flash_read(),
* adding and jffs2_flash_read_end ( ) interface . */
if ( c - > mtd - > point ) {
err = c - > mtd - > point ( c - > mtd , ofs , len , & retlen , & buffer ) ;
if ( ! err & & retlen < tn - > csize ) {
JFFS2_WARNING ( " MTD point returned len too short: %zu instead of %u. \n " , retlen , tn - > csize ) ;
c - > mtd - > unpoint ( c - > mtd , buffer , ofs , len ) ;
} else if ( err )
JFFS2_WARNING ( " MTD point failed: error code %d. \n " , err ) ;
else
pointed = 1 ; /* succefully pointed to device */
}
# endif
if ( ! pointed ) {
buffer = kmalloc ( len , GFP_KERNEL ) ;
if ( unlikely ( ! buffer ) )
return - ENOMEM ;
/* TODO: this is very frequent pattern, make it a separate
* routine */
err = jffs2_flash_read ( c , ofs , len , & retlen , buffer ) ;
if ( err ) {
JFFS2_ERROR ( " can not read %d bytes from 0x%08x, error code: %d. \n " , len , ofs , err ) ;
goto free_out ;
}
if ( retlen ! = len ) {
JFFS2_ERROR ( " short read at %#08x: %zd instead of %d. \n " , ofs , retlen , len ) ;
err = - EIO ;
goto free_out ;
}
}
/* Continue calculating CRC */
crc = crc32 ( tn - > partial_crc , buffer , len ) ;
if ( ! pointed )
kfree ( buffer ) ;
# ifndef __ECOS
else
c - > mtd - > unpoint ( c - > mtd , buffer , ofs , len ) ;
# endif
if ( crc ! = tn - > data_crc ) {
JFFS2_NOTICE ( " wrong data CRC in data node at 0x%08x: read %#08x, calculated %#08x. \n " ,
ofs , tn - > data_crc , crc ) ;
return 1 ;
}
adj_acc :
jeb = & c - > blocks [ ref - > flash_offset / c - > sector_size ] ;
len = ref_totlen ( c , jeb , ref ) ;
/* If it should be REF_NORMAL, it'll get marked as such when
we build the fragtree , shortly . No need to worry about GC
moving it while it ' s marked REF_PRISTINE - - GC won ' t happen
till we ' ve finished checking every inode anyway . */
ref - > flash_offset | = REF_PRISTINE ;
/*
* Mark the node as having been checked and fix the
* accounting accordingly .
*/
spin_lock ( & c - > erase_completion_lock ) ;
jeb - > used_size + = len ;
jeb - > unchecked_size - = len ;
c - > used_size + = len ;
c - > unchecked_size - = len ;
jffs2_dbg_acct_paranoia_check_nolock ( c , jeb ) ;
spin_unlock ( & c - > erase_completion_lock ) ;
return 0 ;
free_out :
if ( ! pointed )
kfree ( buffer ) ;
# ifndef __ECOS
else
c - > mtd - > unpoint ( c - > mtd , buffer , ofs , len ) ;
# endif
return err ;
}
/*
* Helper function for jffs2_add_older_frag_to_fragtree ( ) .
*
* Checks the node if we are in the checking stage .
*/
static int check_tn_node ( struct jffs2_sb_info * c , struct jffs2_tmp_dnode_info * tn )
{
int ret ;
BUG_ON ( ref_obsolete ( tn - > fn - > raw ) ) ;
/* We only check the data CRC of unchecked nodes */
if ( ref_flags ( tn - > fn - > raw ) ! = REF_UNCHECKED )
return 0 ;
dbg_readinode ( " check node %#04x-%#04x, phys offs %#08x \n " ,
tn - > fn - > ofs , tn - > fn - > ofs + tn - > fn - > size , ref_offset ( tn - > fn - > raw ) ) ;
ret = check_node_data ( c , tn ) ;
if ( unlikely ( ret < 0 ) ) {
JFFS2_ERROR ( " check_node_data() returned error: %d. \n " ,
ret ) ;
} else if ( unlikely ( ret > 0 ) ) {
dbg_readinode ( " CRC error, mark it obsolete. \n " ) ;
jffs2_mark_node_obsolete ( c , tn - > fn - > raw ) ;
}
return ret ;
}
static struct jffs2_tmp_dnode_info * jffs2_lookup_tn ( struct rb_root * tn_root , uint32_t offset )
{
struct rb_node * next ;
struct jffs2_tmp_dnode_info * tn = NULL ;
dbg_readinode ( " root %p, offset %d \n " , tn_root , offset ) ;
next = tn_root - > rb_node ;
while ( next ) {
tn = rb_entry ( next , struct jffs2_tmp_dnode_info , rb ) ;
if ( tn - > fn - > ofs < offset )
next = tn - > rb . rb_right ;
else if ( tn - > fn - > ofs > = offset )
next = tn - > rb . rb_left ;
else
break ;
}
return tn ;
}
static void jffs2_kill_tn ( struct jffs2_sb_info * c , struct jffs2_tmp_dnode_info * tn )
{
jffs2_mark_node_obsolete ( c , tn - > fn - > raw ) ;
jffs2_free_full_dnode ( tn - > fn ) ;
jffs2_free_tmp_dnode_info ( tn ) ;
}
/*
* This function is used when we read an inode . Data nodes arrive in
* arbitrary order - - they may be older or newer than the nodes which
* are already in the tree . Where overlaps occur , the older node can
* be discarded as long as the newer passes the CRC check . We don ' t
* bother to keep track of holes in this rbtree , and neither do we deal
* with frags - - we can have multiple entries starting at the same
* offset , and the one with the smallest length will come first in the
* ordering .
*
* Returns 0 if the node was inserted
* 1 if the node is obsolete ( because we can ' t mark it so yet )
* < 0 an if error occurred
2005-04-16 22:20:36 +00:00
*/
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
static int jffs2_add_tn_to_tree ( struct jffs2_sb_info * c ,
struct jffs2_readinode_info * rii ,
struct jffs2_tmp_dnode_info * tn )
{
uint32_t fn_end = tn - > fn - > ofs + tn - > fn - > size ;
struct jffs2_tmp_dnode_info * insert_point = NULL , * this ;
dbg_readinode ( " insert fragment %#04x-%#04x, ver %u \n " , tn - > fn - > ofs , fn_end , tn - > version ) ;
/* If a node has zero dsize, we only have to keep if it if it might be the
node with highest version - - i . e . the one which will end up as f - > metadata .
Note that such nodes won ' t be REF_UNCHECKED since there are no data to
check anyway . */
if ( ! tn - > fn - > size ) {
if ( rii - > mdata_tn ) {
/* We had a candidate mdata node already */
dbg_readinode ( " kill old mdata with ver %d \n " , rii - > mdata_tn - > version ) ;
jffs2_kill_tn ( c , rii - > mdata_tn ) ;
}
rii - > mdata_tn = tn ;
dbg_readinode ( " keep new mdata with ver %d \n " , tn - > version ) ;
return 0 ;
}
/* Find the earliest node which _may_ be relevant to this one */
this = jffs2_lookup_tn ( & rii - > tn_root , tn - > fn - > ofs ) ;
if ( ! this ) {
/* First addition to empty tree. $DEITY how I love the easy cases */
rb_link_node ( & tn - > rb , NULL , & rii - > tn_root . rb_node ) ;
rb_insert_color ( & tn - > rb , & rii - > tn_root ) ;
dbg_readinode ( " keep new frag \n " ) ;
return 0 ;
}
/* If we add a new node it'll be somewhere under here. */
insert_point = this ;
/* If the node is coincident with another at a lower address,
back up until the other node is found . It may be relevant */
while ( tn - > overlapped )
tn = tn_prev ( tn ) ;
dbg_readinode ( " 'this' found %#04x-%#04x (%s) \n " , this - > fn - > ofs , this - > fn - > ofs + this - > fn - > size , this - > fn ? " data " : " hole " ) ;
while ( this ) {
if ( this - > fn - > ofs > fn_end )
break ;
dbg_readinode ( " Ponder this ver %d, 0x%x-0x%x \n " ,
this - > version , this - > fn - > ofs , this - > fn - > size ) ;
if ( this - > version = = tn - > version ) {
/* Version number collision means REF_PRISTINE GC. Accept either of them
as long as the CRC is correct . Check the one we have already . . . */
if ( ! check_tn_node ( c , this ) ) {
/* The one we already had was OK. Keep it and throw away the new one */
dbg_readinode ( " Like old node. Throw away new \n " ) ;
jffs2_kill_tn ( c , tn ) ;
return 0 ;
} else {
/* Who cares if the new one is good; keep it for now anyway. */
rb_replace_node ( & this - > rb , & tn - > rb , & rii - > tn_root ) ;
/* Same overlapping from in front and behind */
tn - > overlapped = this - > overlapped ;
jffs2_kill_tn ( c , this ) ;
dbg_readinode ( " Like new node. Throw away old \n " ) ;
return 0 ;
}
}
if ( this - > version < tn - > version & &
this - > fn - > ofs > = tn - > fn - > ofs & &
this - > fn - > ofs + this - > fn - > size < = fn_end ) {
/* New node entirely overlaps 'this' */
if ( check_tn_node ( c , tn ) ) {
dbg_readinode ( " new node bad CRC \n " ) ;
jffs2_kill_tn ( c , tn ) ;
return 0 ;
}
/* ... and is good. Kill 'this'... */
rb_replace_node ( & this - > rb , & tn - > rb , & rii - > tn_root ) ;
tn - > overlapped = this - > overlapped ;
jffs2_kill_tn ( c , this ) ;
/* ... and any subsequent nodes which are also overlapped */
this = tn_next ( tn ) ;
while ( this & & this - > fn - > ofs + this - > fn - > size < fn_end ) {
struct jffs2_tmp_dnode_info * next = tn_next ( this ) ;
if ( this - > version < tn - > version ) {
tn_erase ( this , & rii - > tn_root ) ;
dbg_readinode ( " Kill overlapped ver %d, 0x%x-0x%x \n " ,
this - > version , this - > fn - > ofs ,
this - > fn - > ofs + this - > fn - > size ) ;
jffs2_kill_tn ( c , this ) ;
}
this = next ;
}
dbg_readinode ( " Done inserting new \n " ) ;
return 0 ;
}
if ( this - > version > tn - > version & &
this - > fn - > ofs < = tn - > fn - > ofs & &
this - > fn - > ofs + this - > fn - > size > = fn_end ) {
/* New node entirely overlapped by 'this' */
if ( ! check_tn_node ( c , this ) ) {
dbg_readinode ( " Good CRC on old node. Kill new \n " ) ;
jffs2_kill_tn ( c , tn ) ;
return 0 ;
}
/* ... but 'this' was bad. Replace it... */
rb_replace_node ( & this - > rb , & tn - > rb , & rii - > tn_root ) ;
dbg_readinode ( " Bad CRC on old overlapping node. Kill it \n " ) ;
jffs2_kill_tn ( c , this ) ;
return 0 ;
}
/* We want to be inserted under the last node which is
either at a lower offset _or_ has a smaller range */
if ( this - > fn - > ofs < tn - > fn - > ofs | |
( this - > fn - > ofs = = tn - > fn - > ofs & &
this - > fn - > size < = tn - > fn - > size ) )
insert_point = this ;
this = tn_next ( this ) ;
}
dbg_readinode ( " insert_point %p, ver %d, 0x%x-0x%x, ov %d \n " ,
insert_point , insert_point - > version , insert_point - > fn - > ofs ,
insert_point - > fn - > ofs + insert_point - > fn - > size ,
insert_point - > overlapped ) ;
/* We neither completely obsoleted nor were completely
obsoleted by an earlier node . Insert under insert_point */
{
struct rb_node * parent = & insert_point - > rb ;
struct rb_node * * link = & parent ;
while ( * link ) {
parent = * link ;
insert_point = rb_entry ( parent , struct jffs2_tmp_dnode_info , rb ) ;
if ( tn - > fn - > ofs > insert_point - > fn - > ofs )
link = & insert_point - > rb . rb_right ;
else if ( tn - > fn - > ofs < insert_point - > fn - > ofs | |
tn - > fn - > size < insert_point - > fn - > size )
link = & insert_point - > rb . rb_left ;
else
link = & insert_point - > rb . rb_right ;
}
rb_link_node ( & tn - > rb , & insert_point - > rb , link ) ;
rb_insert_color ( & tn - > rb , & rii - > tn_root ) ;
}
/* If there's anything behind that overlaps us, note it */
this = tn_prev ( tn ) ;
if ( this ) {
while ( 1 ) {
if ( this - > fn - > ofs + this - > fn - > size > tn - > fn - > ofs ) {
dbg_readinode ( " Node is overlapped by %p (v %d, 0x%x-0x%x) \n " ,
this , this - > version , this - > fn - > ofs ,
this - > fn - > ofs + this - > fn - > size ) ;
tn - > overlapped = 1 ;
break ;
}
if ( ! this - > overlapped )
break ;
this = tn_prev ( this ) ;
}
}
/* If the new node overlaps anything ahead, note it */
this = tn_next ( tn ) ;
while ( this & & this - > fn - > ofs < fn_end ) {
this - > overlapped = 1 ;
dbg_readinode ( " Node ver %d, 0x%x-0x%x is overlapped \n " ,
this - > version , this - > fn - > ofs ,
this - > fn - > ofs + this - > fn - > size ) ;
this = tn_next ( this ) ;
}
return 0 ;
}
/* Trivial function to remove the last node in the tree. Which by definition
has no right - hand - - so can be removed just by making its only child ( if
any ) take its place under its parent . */
static void eat_last ( struct rb_root * root , struct rb_node * node )
2005-04-16 22:20:36 +00:00
{
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
struct rb_node * parent = rb_parent ( node ) ;
struct rb_node * * link ;
/* LAST! */
BUG_ON ( node - > rb_right ) ;
if ( ! parent )
link = & root - > rb_node ;
else if ( node = = parent - > rb_left )
link = & parent - > rb_left ;
else
link = & parent - > rb_right ;
* link = node - > rb_left ;
/* Colour doesn't matter now. Only the parent pointer. */
if ( node - > rb_left )
node - > rb_left - > rb_parent_color = node - > rb_parent_color ;
}
/* We put this in reverse order, so we can just use eat_last */
static void ver_insert ( struct rb_root * ver_root , struct jffs2_tmp_dnode_info * tn )
{
struct rb_node * * link = & ver_root - > rb_node ;
struct rb_node * parent = NULL ;
struct jffs2_tmp_dnode_info * this_tn ;
while ( * link ) {
parent = * link ;
this_tn = rb_entry ( parent , struct jffs2_tmp_dnode_info , rb ) ;
if ( tn - > version > this_tn - > version )
link = & parent - > rb_left ;
2005-07-27 14:46:14 +00:00
else
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
link = & parent - > rb_right ;
2005-08-01 12:05:22 +00:00
}
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
dbg_readinode ( " Link new node at %p (root is %p) \n " , link , ver_root ) ;
rb_link_node ( & tn - > rb , parent , link ) ;
rb_insert_color ( & tn - > rb , ver_root ) ;
}
2005-07-27 14:46:14 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
/* Build final, normal fragtree from tn tree. It doesn't matter which order
we add nodes to the real fragtree , as long as they don ' t overlap . And
having thrown away the majority of overlapped nodes as we went , there
really shouldn ' t be many sets of nodes which do overlap . If we start at
the end , we can use the overlap markers - - we can just eat nodes which
aren ' t overlapped , and when we encounter nodes which _do_ overlap we
sort them all into a temporary tree in version order before replaying them . */
static int jffs2_build_inode_fragtree ( struct jffs2_sb_info * c ,
struct jffs2_inode_info * f ,
struct jffs2_readinode_info * rii )
{
struct jffs2_tmp_dnode_info * pen , * last , * this ;
struct rb_root ver_root = RB_ROOT ;
uint32_t high_ver = 0 ;
if ( rii - > mdata_tn ) {
dbg_readinode ( " potential mdata is ver %d at %p \n " , rii - > mdata_tn - > version , rii - > mdata_tn ) ;
high_ver = rii - > mdata_tn - > version ;
rii - > latest_ref = rii - > mdata_tn - > fn - > raw ;
}
# ifdef JFFS2_DBG_READINODE_MESSAGES
this = tn_last ( & rii - > tn_root ) ;
while ( this ) {
dbg_readinode ( " tn %p ver %d range 0x%x-0x%x ov %d \n " , this , this - > version , this - > fn - > ofs ,
this - > fn - > ofs + this - > fn - > size , this - > overlapped ) ;
this = tn_prev ( this ) ;
}
# endif
pen = tn_last ( & rii - > tn_root ) ;
while ( ( last = pen ) ) {
pen = tn_prev ( last ) ;
eat_last ( & rii - > tn_root , & last - > rb ) ;
ver_insert ( & ver_root , last ) ;
if ( unlikely ( last - > overlapped ) )
continue ;
/* Now we have a bunch of nodes in reverse version
order , in the tree at ver_root . Most of the time ,
there ' ll actually be only one node in the ' tree ' ,
in fact . */
this = tn_last ( & ver_root ) ;
while ( this ) {
struct jffs2_tmp_dnode_info * vers_next ;
int ret ;
vers_next = tn_prev ( this ) ;
eat_last ( & ver_root , & this - > rb ) ;
if ( check_tn_node ( c , this ) ) {
dbg_readinode ( " node ver %x, 0x%x-0x%x failed CRC \n " ,
this - > version , this - > fn - > ofs ,
this - > fn - > ofs + this - > fn - > size ) ;
jffs2_kill_tn ( c , this ) ;
} else {
if ( this - > version > high_ver ) {
/* Note that this is different from the other
highest_version , because this one is only
counting _valid_ nodes which could give the
latest inode metadata */
high_ver = this - > version ;
rii - > latest_ref = this - > fn - > raw ;
}
dbg_readinode ( " Add %p (v %x, 0x%x-0x%x, ov %d) to fragtree \n " ,
this , this - > version , this - > fn - > ofs ,
this - > fn - > ofs + this - > fn - > size , this - > overlapped ) ;
ret = jffs2_add_full_dnode_to_inode ( c , f , this - > fn ) ;
if ( ret ) {
/* Free the nodes in vers_root; let the caller
deal with the rest */
JFFS2_ERROR ( " Add node to tree failed %d \n " , ret ) ;
while ( 1 ) {
vers_next = tn_prev ( this ) ;
if ( check_tn_node ( c , this ) )
jffs2_mark_node_obsolete ( c , this - > fn - > raw ) ;
jffs2_free_full_dnode ( this - > fn ) ;
jffs2_free_tmp_dnode_info ( this ) ;
this = vers_next ;
if ( ! this )
break ;
eat_last ( & ver_root , & vers_next - > rb ) ;
}
return ret ;
}
jffs2_free_tmp_dnode_info ( this ) ;
}
this = vers_next ;
}
}
return 0 ;
2005-07-27 14:46:14 +00:00
}
2005-04-16 22:20:36 +00:00
2005-07-27 14:46:14 +00:00
static void jffs2_free_tmp_dnode_info_list ( struct rb_root * list )
{
struct rb_node * this ;
struct jffs2_tmp_dnode_info * tn ;
this = list - > rb_node ;
/* Now at bottom of tree */
while ( this ) {
if ( this - > rb_left )
this = this - > rb_left ;
else if ( this - > rb_right )
this = this - > rb_right ;
else {
tn = rb_entry ( this , struct jffs2_tmp_dnode_info , rb ) ;
jffs2_free_full_dnode ( tn - > fn ) ;
jffs2_free_tmp_dnode_info ( tn ) ;
2006-04-21 12:17:57 +00:00
this = rb_parent ( this ) ;
2005-07-27 14:46:14 +00:00
if ( ! this )
break ;
if ( this - > rb_left = = & tn - > rb )
this - > rb_left = NULL ;
else if ( this - > rb_right = = & tn - > rb )
this - > rb_right = NULL ;
else BUG ( ) ;
}
}
list - > rb_node = NULL ;
}
2005-04-16 22:20:36 +00:00
2005-07-27 14:46:14 +00:00
static void jffs2_free_full_dirent_list ( struct jffs2_full_dirent * fd )
{
struct jffs2_full_dirent * next ;
2005-07-07 15:45:32 +00:00
2005-07-27 14:46:14 +00:00
while ( fd ) {
next = fd - > next ;
jffs2_free_full_dirent ( fd ) ;
fd = next ;
}
}
2005-04-16 22:20:36 +00:00
2005-07-27 14:46:14 +00:00
/* Returns first valid node after 'ref'. May return 'ref' */
static struct jffs2_raw_node_ref * jffs2_first_valid_node ( struct jffs2_raw_node_ref * ref )
{
while ( ref & & ref - > next_in_ino ) {
if ( ! ref_obsolete ( ref ) )
return ref ;
2005-09-22 11:25:00 +00:00
dbg_noderef ( " node at 0x%08x is obsoleted. Ignoring. \n " , ref_offset ( ref ) ) ;
2005-07-27 14:46:14 +00:00
ref = ref - > next_in_ino ;
}
return NULL ;
}
2005-04-16 22:20:36 +00:00
2005-07-27 14:46:14 +00:00
/*
* Helper function for jffs2_get_inode_nodes ( ) .
* It is called every time an directory entry node is found .
*
* Returns : 0 on succes ;
* 1 if the node should be marked obsolete ;
* negative error code on failure .
*/
2005-08-01 12:05:22 +00:00
static inline int read_direntry ( struct jffs2_sb_info * c , struct jffs2_raw_node_ref * ref ,
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
struct jffs2_raw_dirent * rd , size_t read ,
struct jffs2_readinode_info * rii )
2005-07-27 14:46:14 +00:00
{
struct jffs2_full_dirent * fd ;
2006-06-18 21:44:21 +00:00
uint32_t crc ;
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
/* Obsoleted. This cannot happen, surely? dwmw2 20020308 */
BUG_ON ( ref_obsolete ( ref ) ) ;
2005-11-07 11:16:07 +00:00
2006-06-18 21:44:21 +00:00
crc = crc32 ( 0 , rd , sizeof ( * rd ) - 8 ) ;
if ( unlikely ( crc ! = je32_to_cpu ( rd - > node_crc ) ) ) {
JFFS2_NOTICE ( " header CRC failed on dirent node at %#08x: read %#08x, calculated %#08x \n " ,
ref_offset ( ref ) , je32_to_cpu ( rd - > node_crc ) , crc ) ;
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
jffs2_mark_node_obsolete ( c , ref ) ;
return 0 ;
2005-07-27 14:46:14 +00:00
}
2005-11-07 11:16:07 +00:00
2006-06-18 21:44:21 +00:00
/* If we've never checked the CRCs on this node, check them now */
if ( ref_flags ( ref ) = = REF_UNCHECKED ) {
struct jffs2_eraseblock * jeb ;
int len ;
/* Sanity check */
if ( unlikely ( PAD ( ( rd - > nsize + sizeof ( * rd ) ) ) ! = PAD ( je32_to_cpu ( rd - > totlen ) ) ) ) {
JFFS2_ERROR ( " illegal nsize in node at %#08x: nsize %#02x, totlen %#04x \n " ,
ref_offset ( ref ) , rd - > nsize , je32_to_cpu ( rd - > totlen ) ) ;
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
jffs2_mark_node_obsolete ( c , ref ) ;
return 0 ;
2006-06-18 21:44:21 +00:00
}
jeb = & c - > blocks [ ref - > flash_offset / c - > sector_size ] ;
len = ref_totlen ( c , jeb , ref ) ;
spin_lock ( & c - > erase_completion_lock ) ;
jeb - > used_size + = len ;
jeb - > unchecked_size - = len ;
c - > used_size + = len ;
c - > unchecked_size - = len ;
ref - > flash_offset = ref_offset ( ref ) | REF_PRISTINE ;
spin_unlock ( & c - > erase_completion_lock ) ;
}
2005-07-27 14:46:14 +00:00
fd = jffs2_alloc_full_dirent ( rd - > nsize + 1 ) ;
if ( unlikely ( ! fd ) )
return - ENOMEM ;
2005-04-16 22:20:36 +00:00
2005-07-27 14:46:14 +00:00
fd - > raw = ref ;
fd - > version = je32_to_cpu ( rd - > version ) ;
fd - > ino = je32_to_cpu ( rd - > ino ) ;
fd - > type = rd - > type ;
2005-04-16 22:20:36 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
if ( fd - > version > rii - > highest_version )
rii - > highest_version = fd - > version ;
2005-07-27 14:46:14 +00:00
/* Pick out the mctime of the latest dirent */
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
if ( fd - > version > rii - > mctime_ver & & je32_to_cpu ( rd - > mctime ) ) {
rii - > mctime_ver = fd - > version ;
rii - > latest_mctime = je32_to_cpu ( rd - > mctime ) ;
2005-04-16 22:20:36 +00:00
}
2005-11-07 11:16:07 +00:00
/*
2005-07-27 14:46:14 +00:00
* Copy as much of the name as possible from the raw
* dirent we ' ve already read from the flash .
*/
if ( read > sizeof ( * rd ) )
memcpy ( & fd - > name [ 0 ] , & rd - > name [ 0 ] ,
min_t ( uint32_t , rd - > nsize , ( read - sizeof ( * rd ) ) ) ) ;
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
/* Do we need to copy any more of the name directly from the flash? */
if ( rd - > nsize + sizeof ( * rd ) > read ) {
/* FIXME: point() */
int err ;
int already = read - sizeof ( * rd ) ;
2005-11-07 11:16:07 +00:00
err = jffs2_flash_read ( c , ( ref_offset ( ref ) ) + read ,
2005-07-27 14:46:14 +00:00
rd - > nsize - already , & read , & fd - > name [ already ] ) ;
if ( unlikely ( read ! = rd - > nsize - already ) & & likely ( ! err ) )
return - EIO ;
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
if ( unlikely ( err ) ) {
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " read remainder of name: error %d \n " , err ) ;
2005-07-27 14:46:14 +00:00
jffs2_free_full_dirent ( fd ) ;
return - EIO ;
2005-04-16 22:20:36 +00:00
}
}
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
fd - > nhash = full_name_hash ( fd - > name , rd - > nsize ) ;
fd - > next = NULL ;
fd - > name [ rd - > nsize ] = ' \0 ' ;
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
/*
* Wheee . We now have a complete jffs2_full_dirent structure , with
2005-11-07 11:16:07 +00:00
* the name in it and everything . Link it into the list
2005-07-27 14:46:14 +00:00
*/
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
jffs2_add_fd_to_list ( c , fd , & rii - > fds ) ;
2005-07-27 14:46:14 +00:00
2005-04-16 22:20:36 +00:00
return 0 ;
}
2005-07-27 14:46:14 +00:00
/*
* Helper function for jffs2_get_inode_nodes ( ) .
* It is called every time an inode node is found .
*
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
* Returns : 0 on success ;
2005-07-27 14:46:14 +00:00
* 1 if the node should be marked obsolete ;
* negative error code on failure .
*/
2005-08-01 12:05:22 +00:00
static inline int read_dnode ( struct jffs2_sb_info * c , struct jffs2_raw_node_ref * ref ,
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
struct jffs2_raw_inode * rd , int rdlen ,
struct jffs2_readinode_info * rii )
2005-04-16 22:20:36 +00:00
{
2005-07-27 14:46:14 +00:00
struct jffs2_tmp_dnode_info * tn ;
2005-08-01 12:05:22 +00:00
uint32_t len , csize ;
int ret = 1 ;
2006-06-18 21:44:21 +00:00
uint32_t crc ;
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
/* Obsoleted. This cannot happen, surely? dwmw2 20020308 */
BUG_ON ( ref_obsolete ( ref ) ) ;
2006-06-18 21:44:21 +00:00
crc = crc32 ( 0 , rd , sizeof ( * rd ) - 8 ) ;
if ( unlikely ( crc ! = je32_to_cpu ( rd - > node_crc ) ) ) {
JFFS2_NOTICE ( " node CRC failed on dnode at %#08x: read %#08x, calculated %#08x \n " ,
ref_offset ( ref ) , je32_to_cpu ( rd - > node_crc ) , crc ) ;
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
jffs2_mark_node_obsolete ( c , ref ) ;
return 0 ;
2006-06-18 21:44:21 +00:00
}
2005-08-01 12:05:22 +00:00
tn = jffs2_alloc_tmp_dnode_info ( ) ;
if ( ! tn ) {
2006-04-12 00:12:10 +00:00
JFFS2_ERROR ( " failed to allocate tn (%zu bytes). \n " , sizeof ( * tn ) ) ;
2005-08-01 12:05:22 +00:00
return - ENOMEM ;
}
tn - > partial_crc = 0 ;
csize = je32_to_cpu ( rd - > csize ) ;
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
/* If we've never checked the CRCs on this node, check them now */
if ( ref_flags ( ref ) = = REF_UNCHECKED ) {
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
/* Sanity checks */
if ( unlikely ( je32_to_cpu ( rd - > offset ) > je32_to_cpu ( rd - > isize ) ) | |
unlikely ( PAD ( je32_to_cpu ( rd - > csize ) + sizeof ( * rd ) ) ! = PAD ( je32_to_cpu ( rd - > totlen ) ) ) ) {
2005-07-28 14:46:43 +00:00
JFFS2_WARNING ( " inode node header CRC is corrupted at %#08x \n " , ref_offset ( ref ) ) ;
2005-07-30 15:29:30 +00:00
jffs2_dbg_dump_node ( c , ref_offset ( ref ) ) ;
2005-08-01 12:05:22 +00:00
goto free_out ;
2005-04-16 22:20:36 +00:00
}
2005-08-01 12:05:22 +00:00
if ( jffs2_is_writebuffered ( c ) & & csize ! = 0 ) {
/* At this point we are supposed to check the data CRC
* of our unchecked node . But thus far , we do not
* know whether the node is valid or obsolete . To
* figure this out , we need to walk all the nodes of
* the inode and build the inode fragtree . We don ' t
* want to spend time checking data of nodes which may
* later be found to be obsolete . So we put off the full
* data CRC checking until we have read all the inode
* nodes and have started building the fragtree .
*
* The fragtree is being built starting with nodes
* having the highest version number , so we ' ll be able
* to detect whether a node is valid ( i . e . , it is not
* overlapped by a node with higher version ) or not .
* And we ' ll be able to check only those nodes , which
* are not obsolete .
*
* Of course , this optimization only makes sense in case
* of NAND flashes ( or other flashes whith
* ! jffs2_can_mark_obsolete ( ) ) , since on NOR flashes
* nodes are marked obsolete physically .
*
* Since NAND flashes ( or other flashes with
* jffs2_is_writebuffered ( c ) ) are anyway read by
* fractions of c - > wbuf_pagesize , and we have just read
* the node header , it is likely that the starting part
* of the node data is also read when we read the
* header . So we don ' t mind to check the CRC of the
* starting part of the data of the node now , and check
* the second part later ( in jffs2_check_node_data ( ) ) .
* Of course , we will not need to re - read and re - check
* the NAND page which we have just read . This is why we
* read the whole NAND page at jffs2_get_inode_nodes ( ) ,
* while we needed only the node header .
*/
unsigned char * buf ;
/* 'buf' will point to the start of data */
buf = ( unsigned char * ) rd + sizeof ( * rd ) ;
/* len will be the read data length */
len = min_t ( uint32_t , rdlen - sizeof ( * rd ) , csize ) ;
2005-08-17 14:57:43 +00:00
tn - > partial_crc = crc32 ( 0 , buf , len ) ;
2005-09-22 11:25:00 +00:00
dbg_readinode ( " Calculates CRC (%#08x) for %d bytes, csize %d \n " , tn - > partial_crc , len , csize ) ;
2005-08-01 12:05:22 +00:00
/* If we actually calculated the whole data CRC
* and it is wrong , drop the node . */
2005-08-04 11:40:02 +00:00
if ( len > = csize & & unlikely ( tn - > partial_crc ! = je32_to_cpu ( rd - > data_crc ) ) ) {
2005-08-03 09:26:50 +00:00
JFFS2_NOTICE ( " wrong data CRC in data node at 0x%08x: read %#08x, calculated %#08x. \n " ,
ref_offset ( ref ) , tn - > partial_crc , je32_to_cpu ( rd - > data_crc ) ) ;
2005-08-01 12:05:22 +00:00
goto free_out ;
2005-08-03 09:26:50 +00:00
}
2005-04-16 22:20:36 +00:00
2005-08-01 12:05:22 +00:00
} else if ( csize = = 0 ) {
/*
* We checked the header CRC . If the node has no data , adjust
* the space accounting now . For other nodes this will be done
* later either when the node is marked obsolete or when its
* data is checked .
*/
struct jffs2_eraseblock * jeb ;
2005-09-22 11:25:00 +00:00
dbg_readinode ( " the node has no data. \n " ) ;
2005-08-01 12:05:22 +00:00
jeb = & c - > blocks [ ref - > flash_offset / c - > sector_size ] ;
len = ref_totlen ( c , jeb , ref ) ;
spin_lock ( & c - > erase_completion_lock ) ;
jeb - > used_size + = len ;
jeb - > unchecked_size - = len ;
c - > used_size + = len ;
c - > unchecked_size - = len ;
2005-07-27 14:46:14 +00:00
ref - > flash_offset = ref_offset ( ref ) | REF_NORMAL ;
2005-08-01 12:05:22 +00:00
spin_unlock ( & c - > erase_completion_lock ) ;
2005-04-16 22:20:36 +00:00
}
}
2005-07-27 14:46:14 +00:00
tn - > fn = jffs2_alloc_full_dnode ( ) ;
if ( ! tn - > fn ) {
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " alloc fn failed \n " ) ;
2005-08-01 12:05:22 +00:00
ret = - ENOMEM ;
goto free_out ;
2005-07-27 14:46:14 +00:00
}
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
tn - > version = je32_to_cpu ( rd - > version ) ;
tn - > fn - > ofs = je32_to_cpu ( rd - > offset ) ;
2005-08-01 12:05:22 +00:00
tn - > data_crc = je32_to_cpu ( rd - > data_crc ) ;
tn - > csize = csize ;
2005-07-27 14:46:14 +00:00
tn - > fn - > raw = ref ;
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
tn - > overlapped = 0 ;
if ( tn - > version > rii - > highest_version )
rii - > highest_version = tn - > version ;
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
/* There was a bug where we wrote hole nodes out with
csize / dsize swapped . Deal with it */
2005-08-01 12:05:22 +00:00
if ( rd - > compr = = JFFS2_COMPR_ZERO & & ! je32_to_cpu ( rd - > dsize ) & & csize )
tn - > fn - > size = csize ;
2005-07-27 14:46:14 +00:00
else // normal case...
tn - > fn - > size = je32_to_cpu ( rd - > dsize ) ;
2005-09-22 11:25:00 +00:00
dbg_readinode ( " dnode @%08x: ver %u, offset %#04x, dsize %#04x, csize %#04x \n " ,
2005-08-17 14:57:43 +00:00
ref_offset ( ref ) , je32_to_cpu ( rd - > version ) , je32_to_cpu ( rd - > offset ) , je32_to_cpu ( rd - > dsize ) , csize ) ;
2005-11-07 11:16:07 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
ret = jffs2_add_tn_to_tree ( c , rii , tn ) ;
2005-04-16 22:20:36 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
if ( ret ) {
jffs2_free_full_dnode ( tn - > fn ) ;
free_out :
jffs2_free_tmp_dnode_info ( tn ) ;
return ret ;
}
# ifdef JFFS2_DBG_READINODE_MESSAGES
dbg_readinode ( " After adding ver %d: \n " , tn - > version ) ;
tn = tn_first ( & rii - > tn_root ) ;
while ( tn ) {
dbg_readinode ( " %p: v %d r 0x%x-0x%x ov %d \n " ,
tn , tn - > version , tn - > fn - > ofs ,
tn - > fn - > ofs + tn - > fn - > size , tn - > overlapped ) ;
tn = tn_next ( tn ) ;
}
# endif
2005-04-16 22:20:36 +00:00
return 0 ;
}
2005-07-27 14:46:14 +00:00
/*
* Helper function for jffs2_get_inode_nodes ( ) .
* It is called every time an unknown node is found .
*
2006-06-17 23:05:26 +00:00
* Returns : 0 on success ;
2005-07-27 14:46:14 +00:00
* 1 if the node should be marked obsolete ;
* negative error code on failure .
*/
2005-08-01 12:05:22 +00:00
static inline int read_unknown ( struct jffs2_sb_info * c , struct jffs2_raw_node_ref * ref , struct jffs2_unknown_node * un )
2005-04-16 22:20:36 +00:00
{
2005-07-27 14:46:14 +00:00
/* We don't mark unknown nodes as REF_UNCHECKED */
2007-03-09 11:44:00 +00:00
if ( ref_flags ( ref ) = = REF_UNCHECKED ) {
JFFS2_ERROR ( " REF_UNCHECKED but unknown node at %#08x \n " ,
ref_offset ( ref ) ) ;
JFFS2_ERROR ( " Node is {%04x,%04x,%08x,%08x}. Please report this error. \n " ,
je16_to_cpu ( un - > magic ) , je16_to_cpu ( un - > nodetype ) ,
je32_to_cpu ( un - > totlen ) , je32_to_cpu ( un - > hdr_crc ) ) ;
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
jffs2_mark_node_obsolete ( c , ref ) ;
return 0 ;
2007-03-09 11:44:00 +00:00
}
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
un - > nodetype = cpu_to_je16 ( JFFS2_NODE_ACCURATE | je16_to_cpu ( un - > nodetype ) ) ;
2005-04-16 22:20:36 +00:00
2006-06-17 23:05:26 +00:00
switch ( je16_to_cpu ( un - > nodetype ) & JFFS2_COMPAT_MASK ) {
case JFFS2_FEATURE_INCOMPAT :
JFFS2_ERROR ( " unknown INCOMPAT nodetype %#04X at %#08x \n " ,
je16_to_cpu ( un - > nodetype ) , ref_offset ( ref ) ) ;
/* EEP */
BUG ( ) ;
break ;
case JFFS2_FEATURE_ROCOMPAT :
JFFS2_ERROR ( " unknown ROCOMPAT nodetype %#04X at %#08x \n " ,
je16_to_cpu ( un - > nodetype ) , ref_offset ( ref ) ) ;
BUG_ON ( ! ( c - > flags & JFFS2_SB_FLAG_RO ) ) ;
break ;
case JFFS2_FEATURE_RWCOMPAT_COPY :
JFFS2_NOTICE ( " unknown RWCOMPAT_COPY nodetype %#04X at %#08x \n " ,
je16_to_cpu ( un - > nodetype ) , ref_offset ( ref ) ) ;
break ;
case JFFS2_FEATURE_RWCOMPAT_DELETE :
JFFS2_NOTICE ( " unknown RWCOMPAT_DELETE nodetype %#04X at %#08x \n " ,
je16_to_cpu ( un - > nodetype ) , ref_offset ( ref ) ) ;
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
jffs2_mark_node_obsolete ( c , ref ) ;
return 0 ;
2005-04-16 22:20:36 +00:00
}
2005-07-27 14:46:14 +00:00
return 0 ;
2005-04-16 22:20:36 +00:00
}
2005-08-01 12:05:22 +00:00
/*
* Helper function for jffs2_get_inode_nodes ( ) .
* The function detects whether more data should be read and reads it if yes .
*
* Returns : 0 on succes ;
* negative error code on failure .
*/
static int read_more ( struct jffs2_sb_info * c , struct jffs2_raw_node_ref * ref ,
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
int needed_len , int * rdlen , unsigned char * buf )
2005-08-01 12:05:22 +00:00
{
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
int err , to_read = needed_len - * rdlen ;
2005-08-01 12:05:22 +00:00
size_t retlen ;
uint32_t offs ;
if ( jffs2_is_writebuffered ( c ) ) {
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
int rem = to_read % c - > wbuf_pagesize ;
2005-08-01 12:05:22 +00:00
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
if ( rem )
to_read + = c - > wbuf_pagesize - rem ;
}
2005-08-01 12:05:22 +00:00
/* We need to read more data */
offs = ref_offset ( ref ) + * rdlen ;
2005-11-07 11:16:07 +00:00
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
dbg_readinode ( " read more %d bytes \n " , to_read ) ;
2005-08-01 12:05:22 +00:00
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
err = jffs2_flash_read ( c , offs , to_read , & retlen , buf + * rdlen ) ;
2005-08-01 12:05:22 +00:00
if ( err ) {
JFFS2_ERROR ( " can not read %d bytes from 0x%08x, "
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
" error code: %d. \n " , to_read , offs , err ) ;
2005-08-01 12:05:22 +00:00
return err ;
}
2005-11-07 11:16:07 +00:00
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
if ( retlen < to_read ) {
2006-04-12 00:12:10 +00:00
JFFS2_ERROR ( " short read at %#08x: %zu instead of %d. \n " ,
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
offs , retlen , to_read ) ;
2005-08-01 12:05:22 +00:00
return - EIO ;
}
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
* rdlen + = to_read ;
2005-08-01 12:05:22 +00:00
return 0 ;
}
2005-07-27 14:46:14 +00:00
/* Get tmp_dnode_info and full_dirent for all non-obsolete nodes associated
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
with this ino . Perform a preliminary ordering on data nodes , throwing away
those which are completely obsoleted by newer ones . The naïve approach we
use to take of just returning them _all_ in version order will cause us to
run out of memory in certain degenerate cases . */
2005-07-27 14:46:14 +00:00
static int jffs2_get_inode_nodes ( struct jffs2_sb_info * c , struct jffs2_inode_info * f ,
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
struct jffs2_readinode_info * rii )
2005-04-16 22:20:36 +00:00
{
2005-07-27 14:46:14 +00:00
struct jffs2_raw_node_ref * ref , * valid_ref ;
2005-08-01 12:05:22 +00:00
unsigned char * buf = NULL ;
union jffs2_node_union * node ;
2005-07-27 14:46:14 +00:00
size_t retlen ;
2005-08-01 12:05:22 +00:00
int len , err ;
2005-04-16 22:20:36 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
rii - > mctime_ver = 0 ;
2005-11-07 11:16:07 +00:00
2005-09-22 11:25:00 +00:00
dbg_readinode ( " ino #%u \n " , f - > inocache - > ino ) ;
2005-04-16 22:20:36 +00:00
2005-08-01 12:05:22 +00:00
/* FIXME: in case of NOR and available ->point() this
* needs to be fixed . */
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
len = sizeof ( union jffs2_node_union ) + c - > wbuf_pagesize ;
2005-08-01 12:05:22 +00:00
buf = kmalloc ( len , GFP_KERNEL ) ;
if ( ! buf )
return - ENOMEM ;
2005-11-07 11:16:07 +00:00
2005-08-01 12:05:22 +00:00
spin_lock ( & c - > erase_completion_lock ) ;
2005-07-27 14:46:14 +00:00
valid_ref = jffs2_first_valid_node ( f - > inocache - > nodes ) ;
2005-08-01 12:05:22 +00:00
if ( ! valid_ref & & f - > inocache - > ino ! = 1 )
JFFS2_WARNING ( " Eep. No valid nodes for ino #%u. \n " , f - > inocache - > ino ) ;
2005-07-27 14:46:14 +00:00
while ( valid_ref ) {
/* We can hold a pointer to a non-obsolete node without the spinlock,
but _obsolete_ nodes may disappear at any time , if the block
they ' re in gets erased . So if we mark ' ref ' obsolete while we ' re
not holding the lock , it can go away immediately . For that reason ,
we find the next valid node first , before processing ' ref ' .
*/
ref = valid_ref ;
valid_ref = jffs2_first_valid_node ( ref - > next_in_ino ) ;
spin_unlock ( & c - > erase_completion_lock ) ;
cond_resched ( ) ;
2005-08-01 12:05:22 +00:00
/*
* At this point we don ' t know the type of the node we ' re going
* to read , so we do not know the size of its header . In order
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
* to minimize the amount of flash IO we assume the header is
* of size = JFFS2_MIN_NODE_HEADER .
2005-08-01 12:05:22 +00:00
*/
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
len = JFFS2_MIN_NODE_HEADER ;
2005-08-01 12:05:22 +00:00
if ( jffs2_is_writebuffered ( c ) ) {
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
int end , rem ;
2005-11-07 11:16:07 +00:00
/*
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
* We are about to read JFFS2_MIN_NODE_HEADER bytes ,
* but this flash has some minimal I / O unit . It is
* possible that we ' ll need to read more soon , so read
* up to the next min . I / O unit , in order not to
* re - read the same min . I / O unit twice .
2005-08-01 12:05:22 +00:00
*/
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
end = ref_offset ( ref ) + len ;
rem = end % c - > wbuf_pagesize ;
if ( rem )
end + = c - > wbuf_pagesize - rem ;
len = end - ref_offset ( ref ) ;
2005-08-01 12:05:22 +00:00
}
2005-09-22 11:25:00 +00:00
dbg_readinode ( " read %d bytes at %#08x(%d). \n " , len , ref_offset ( ref ) , ref_flags ( ref ) ) ;
2005-08-01 12:05:22 +00:00
2005-07-27 14:46:14 +00:00
/* FIXME: point() */
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
err = jffs2_flash_read ( c , ref_offset ( ref ) , len , & retlen , buf ) ;
2005-07-27 14:46:14 +00:00
if ( err ) {
2005-08-01 12:05:22 +00:00
JFFS2_ERROR ( " can not read %d bytes from 0x%08x, " " error code: %d. \n " , len , ref_offset ( ref ) , err ) ;
goto free_out ;
}
2005-11-07 11:16:07 +00:00
2005-08-01 12:05:22 +00:00
if ( retlen < len ) {
2006-04-12 00:12:10 +00:00
JFFS2_ERROR ( " short read at %#08x: %zu instead of %d. \n " , ref_offset ( ref ) , retlen , len ) ;
2005-08-01 12:05:22 +00:00
err = - EIO ;
2005-07-27 14:46:14 +00:00
goto free_out ;
}
2005-11-07 11:16:07 +00:00
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
node = ( union jffs2_node_union * ) buf ;
2005-11-07 11:16:07 +00:00
2006-06-17 23:05:26 +00:00
/* No need to mask in the valid bit; it shouldn't be invalid */
if ( je32_to_cpu ( node - > u . hdr_crc ) ! = crc32 ( 0 , node , sizeof ( node - > u ) - 4 ) ) {
JFFS2_NOTICE ( " Node header CRC failed at %#08x. {%04x,%04x,%08x,%08x} \n " ,
ref_offset ( ref ) , je16_to_cpu ( node - > u . magic ) ,
je16_to_cpu ( node - > u . nodetype ) ,
je32_to_cpu ( node - > u . totlen ) ,
je32_to_cpu ( node - > u . hdr_crc ) ) ;
jffs2_dbg_dump_node ( c , ref_offset ( ref ) ) ;
jffs2_mark_node_obsolete ( c , ref ) ;
goto cont ;
}
2007-03-10 16:08:44 +00:00
if ( je16_to_cpu ( node - > u . magic ) ! = JFFS2_MAGIC_BITMASK ) {
/* Not a JFFS2 node, whinge and move on */
JFFS2_NOTICE ( " Wrong magic bitmask 0x%04x in node header at %#08x. \n " ,
je16_to_cpu ( node - > u . magic ) , ref_offset ( ref ) ) ;
2007-03-09 11:44:00 +00:00
jffs2_mark_node_obsolete ( c , ref ) ;
goto cont ;
}
2006-06-17 23:05:26 +00:00
2005-08-01 12:05:22 +00:00
switch ( je16_to_cpu ( node - > u . nodetype ) ) {
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
case JFFS2_NODETYPE_DIRENT :
2005-08-01 12:05:22 +00:00
if ( JFFS2_MIN_NODE_HEADER < sizeof ( struct jffs2_raw_dirent ) ) {
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
err = read_more ( c , ref , sizeof ( struct jffs2_raw_dirent ) , & len , buf ) ;
2005-08-01 12:05:22 +00:00
if ( unlikely ( err ) )
goto free_out ;
}
2005-11-07 11:16:07 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
err = read_direntry ( c , ref , & node - > d , retlen , rii ) ;
if ( unlikely ( err ) )
2005-07-27 14:46:14 +00:00
goto free_out ;
2005-11-07 11:16:07 +00:00
2005-04-16 22:20:36 +00:00
break ;
2005-07-27 14:46:14 +00:00
case JFFS2_NODETYPE_INODE :
2005-11-07 11:16:07 +00:00
2005-08-01 12:05:22 +00:00
if ( JFFS2_MIN_NODE_HEADER < sizeof ( struct jffs2_raw_inode ) ) {
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
err = read_more ( c , ref , sizeof ( struct jffs2_raw_inode ) , & len , buf ) ;
2005-08-01 12:05:22 +00:00
if ( unlikely ( err ) )
goto free_out ;
2005-07-27 14:46:14 +00:00
}
2005-04-16 22:20:36 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
err = read_dnode ( c , ref , & node - > i , len , rii ) ;
if ( unlikely ( err ) )
2005-07-27 14:46:14 +00:00
goto free_out ;
2005-04-16 22:20:36 +00:00
2005-07-27 14:46:14 +00:00
break ;
2005-04-16 22:20:36 +00:00
2005-07-27 14:46:14 +00:00
default :
2005-08-01 12:05:22 +00:00
if ( JFFS2_MIN_NODE_HEADER < sizeof ( struct jffs2_unknown_node ) ) {
[JFFS2] fix buffer sise calculations in jffs2_get_inode_nodes()
In read inode we have an optimization which prevents one
min. I/O unit (e.g. NAND page) to be read more then once.
Namely, at the beginning we do not know which node type we read,
so we read so we assume we read the directory entry, because it
has the smallest node header. When we read it, we read up to the
next min. I/O unit, just because if later we'll need to read more,
we already have this data.
If it turns out to be that the node is not directory entry, and
we need more data, and we did not read it because it sits in the
next min. I/O unit, we read the whole next (or several next)
min. I/O unit(s). And if it happens to be that we read a data node,
and we've read part of its data, we calculate partial CRC.
So if later we need to check data CRC, we'll only read the rest
of the data from further min. I/O units and continue CRC checking.
This code was a bit messy and buggy. The bug was that it assumed
relatively large min. I/O unit, so that the largest node header
could overlap only one min. I/O unit boundary.
This parch clean-ups the code a bit and fixes this bug.
The patch was not tested on flash with small min. I/O unit, like
NOR-ECC, nut it was tested on NAND with 512 bytes NAND page, so
it at least does not break NAND. It was also tested with mtdram
so it should not break NOR.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-04 10:59:11 +00:00
err = read_more ( c , ref , sizeof ( struct jffs2_unknown_node ) , & len , buf ) ;
2005-08-01 12:05:22 +00:00
if ( unlikely ( err ) )
goto free_out ;
2005-07-27 14:46:14 +00:00
}
2005-11-07 11:16:07 +00:00
2005-08-01 12:05:22 +00:00
err = read_unknown ( c , ref , & node - > u ) ;
2005-07-27 14:46:14 +00:00
if ( err = = 1 ) {
jffs2_mark_node_obsolete ( c , ref ) ;
break ;
} else if ( unlikely ( err ) )
goto free_out ;
}
2006-06-17 23:05:26 +00:00
cont :
2005-07-27 14:46:14 +00:00
spin_lock ( & c - > erase_completion_lock ) ;
2005-04-16 22:20:36 +00:00
}
2005-08-01 12:05:22 +00:00
2005-07-27 14:46:14 +00:00
spin_unlock ( & c - > erase_completion_lock ) ;
2005-08-01 12:05:22 +00:00
kfree ( buf ) ;
2005-07-27 14:46:14 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
f - > highest_version = rii - > highest_version ;
2005-09-22 11:25:00 +00:00
dbg_readinode ( " nodes of inode #%u were read, the highest version is %u, latest_mctime %u, mctime_ver %u. \n " ,
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
f - > inocache - > ino , rii - > highest_version , rii - > latest_mctime ,
rii - > mctime_ver ) ;
2005-07-27 14:46:14 +00:00
return 0 ;
free_out :
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
jffs2_free_tmp_dnode_info_list ( & rii - > tn_root ) ;
jffs2_free_full_dirent_list ( rii - > fds ) ;
rii - > fds = NULL ;
2005-08-01 12:05:22 +00:00
kfree ( buf ) ;
2005-07-27 14:46:14 +00:00
return err ;
2005-04-16 22:20:36 +00:00
}
2005-11-07 11:16:07 +00:00
static int jffs2_do_read_inode_internal ( struct jffs2_sb_info * c ,
2005-04-16 22:20:36 +00:00
struct jffs2_inode_info * f ,
struct jffs2_raw_inode * latest_node )
{
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
struct jffs2_readinode_info rii ;
2005-04-16 22:20:36 +00:00
uint32_t crc ;
size_t retlen ;
int ret ;
2005-09-22 11:25:00 +00:00
dbg_readinode ( " ino #%u nlink is %d \n " , f - > inocache - > ino , f - > inocache - > nlink ) ;
2005-04-16 22:20:36 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
memset ( & rii , 0 , sizeof ( rii ) ) ;
2005-04-16 22:20:36 +00:00
/* Grab all nodes relevant to this ino */
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
ret = jffs2_get_inode_nodes ( c , f , & rii ) ;
2005-04-16 22:20:36 +00:00
if ( ret ) {
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " cannot read nodes for ino %u, returned error is %d \n " , f - > inocache - > ino , ret ) ;
2005-04-16 22:20:36 +00:00
if ( f - > inocache - > state = = INO_STATE_READING )
jffs2_set_inocache_state ( c , f - > inocache , INO_STATE_CHECKEDABSENT ) ;
return ret ;
}
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
ret = jffs2_build_inode_fragtree ( c , f , & rii ) ;
if ( ret ) {
JFFS2_ERROR ( " Failed to build final fragtree for inode #%u: error %d \n " ,
f - > inocache - > ino , ret ) ;
if ( f - > inocache - > state = = INO_STATE_READING )
jffs2_set_inocache_state ( c , f - > inocache , INO_STATE_CHECKEDABSENT ) ;
jffs2_free_tmp_dnode_info_list ( & rii . tn_root ) ;
/* FIXME: We could at least crc-check them all */
if ( rii . mdata_tn ) {
jffs2_free_full_dnode ( rii . mdata_tn - > fn ) ;
jffs2_free_tmp_dnode_info ( rii . mdata_tn ) ;
rii . mdata_tn = NULL ;
}
return ret ;
}
2005-07-05 21:03:10 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
if ( rii . mdata_tn ) {
if ( rii . mdata_tn - > fn - > raw = = rii . latest_ref ) {
f - > metadata = rii . mdata_tn - > fn ;
jffs2_free_tmp_dnode_info ( rii . mdata_tn ) ;
} else {
jffs2_kill_tn ( c , rii . mdata_tn ) ;
2005-08-01 12:05:22 +00:00
}
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
rii . mdata_tn = NULL ;
2005-04-16 22:20:36 +00:00
}
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
f - > dents = rii . fds ;
jffs2_dbg_fragtree_paranoia_check_nolock ( f ) ;
2005-08-01 12:05:22 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
if ( unlikely ( ! rii . latest_ref ) ) {
2005-04-16 22:20:36 +00:00
/* No data nodes for this inode. */
if ( f - > inocache - > ino ! = 1 ) {
2005-07-28 14:46:43 +00:00
JFFS2_WARNING ( " no data nodes found for ino #%u \n " , f - > inocache - > ino ) ;
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
if ( ! rii . fds ) {
2005-04-16 22:20:36 +00:00
if ( f - > inocache - > state = = INO_STATE_READING )
jffs2_set_inocache_state ( c , f - > inocache , INO_STATE_CHECKEDABSENT ) ;
return - EIO ;
}
2005-07-28 14:46:43 +00:00
JFFS2_NOTICE ( " but it has children so we fake some modes for it \n " ) ;
2005-04-16 22:20:36 +00:00
}
latest_node - > mode = cpu_to_jemode ( S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO ) ;
latest_node - > version = cpu_to_je32 ( 0 ) ;
latest_node - > atime = latest_node - > ctime = latest_node - > mtime = cpu_to_je32 ( 0 ) ;
latest_node - > isize = cpu_to_je32 ( 0 ) ;
latest_node - > gid = cpu_to_je16 ( 0 ) ;
latest_node - > uid = cpu_to_je16 ( 0 ) ;
if ( f - > inocache - > state = = INO_STATE_READING )
jffs2_set_inocache_state ( c , f - > inocache , INO_STATE_PRESENT ) ;
return 0 ;
}
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
ret = jffs2_flash_read ( c , ref_offset ( rii . latest_ref ) , sizeof ( * latest_node ) , & retlen , ( void * ) latest_node ) ;
2005-04-16 22:20:36 +00:00
if ( ret | | retlen ! = sizeof ( * latest_node ) ) {
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " failed to read from flash: error %d, %zd of %zd bytes read \n " ,
ret , retlen , sizeof ( * latest_node ) ) ;
2005-04-16 22:20:36 +00:00
/* FIXME: If this fails, there seems to be a memory leak. Find it. */
up ( & f - > sem ) ;
jffs2_do_clear_inode ( c , f ) ;
return ret ? ret : - EIO ;
}
crc = crc32 ( 0 , latest_node , sizeof ( * latest_node ) - 8 ) ;
if ( crc ! = je32_to_cpu ( latest_node - > node_crc ) ) {
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " CRC failed for read_inode of inode %u at physical location 0x%x \n " ,
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
f - > inocache - > ino , ref_offset ( rii . latest_ref ) ) ;
2005-04-16 22:20:36 +00:00
up ( & f - > sem ) ;
jffs2_do_clear_inode ( c , f ) ;
return - EIO ;
}
switch ( jemode_to_cpu ( latest_node - > mode ) & S_IFMT ) {
case S_IFDIR :
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
if ( rii . mctime_ver > je32_to_cpu ( latest_node - > version ) ) {
2005-04-16 22:20:36 +00:00
/* The times in the latest_node are actually older than
mctime in the latest dirent . Cheat . */
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
latest_node - > ctime = latest_node - > mtime = cpu_to_je32 ( rii . latest_mctime ) ;
2005-04-16 22:20:36 +00:00
}
break ;
2005-11-07 11:16:07 +00:00
2005-04-16 22:20:36 +00:00
case S_IFREG :
/* If it was a regular file, truncate it to the latest node's isize */
2005-07-24 15:29:59 +00:00
jffs2_truncate_fragtree ( c , & f - > fragtree , je32_to_cpu ( latest_node - > isize ) ) ;
2005-04-16 22:20:36 +00:00
break ;
case S_IFLNK :
/* Hack to work around broken isize in old symlink code.
Remove this when dwmw2 comes to his senses and stops
symlinks from being an entirely gratuitous special
case . */
if ( ! je32_to_cpu ( latest_node - > isize ) )
latest_node - > isize = latest_node - > dsize ;
2005-03-01 10:50:52 +00:00
if ( f - > inocache - > state ! = INO_STATE_CHECKING ) {
/* Symlink's inode data is the target path. Read it and
2005-07-17 11:13:51 +00:00
* keep in RAM to facilitate quick follow symlink
* operation . */
f - > target = kmalloc ( je32_to_cpu ( latest_node - > csize ) + 1 , GFP_KERNEL ) ;
if ( ! f - > target ) {
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " can't allocate %d bytes of memory for the symlink target path cache \n " , je32_to_cpu ( latest_node - > csize ) ) ;
2005-03-01 10:50:52 +00:00
up ( & f - > sem ) ;
jffs2_do_clear_inode ( c , f ) ;
return - ENOMEM ;
}
2005-11-07 11:16:07 +00:00
[JFFS2] Improve read_inode memory usage, v2.
We originally used to read every node and allocate a jffs2_tmp_dnode_info
structure for each, before processing them in (reverse) version order
and discarding the ones which are obsoleted by later nodes.
With huge logfiles, this behaviour caused memory problems. For example, a
file involved in OLPC trac #1292 has 1822391 nodes, and would cause the XO
machine to run out of memory during the first stage of read_inode().
Instead of just inserting nodes into a tree in version order as we find
them, we now put them into a tree in order of their offset within the
file, which allows us to immediately discard nodes which are completely
obsoleted.
We don't use a full tree with 'fragments' pointing to the real data
structure, as we do in the normal fragtree. We sort only on the start
address, and add an 'overlapped' flag to the tmp_dnode_info to indicate
that the node in question is (partially) overlapped by another.
When the scan is complete, we start at the end of the file, adding each
node to a real fragtree as before. Where the node is non-overlapped, we
just add it (it doesn't matter that it's not the latest version; there is
no overlap). When the node at the end of the tree _is_ overlapped, we sort
it and all its overlapping nodes into version order and then add them to
the fragtree in that order.
This 'early discard' reduces the peak allocation of tmp_dnode_info
structures from 1.8M to a mere 62872 (3.5%) in the degenerate case
referenced above.
This version of the patch also correctly rememembers the highest node
version# seen for an inode when it's scanned.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
2007-04-25 02:23:42 +00:00
ret = jffs2_flash_read ( c , ref_offset ( rii . latest_ref ) + sizeof ( * latest_node ) ,
2005-07-17 11:13:51 +00:00
je32_to_cpu ( latest_node - > csize ) , & retlen , ( char * ) f - > target ) ;
2005-11-07 11:16:07 +00:00
2005-03-01 10:50:52 +00:00
if ( ret | | retlen ! = je32_to_cpu ( latest_node - > csize ) ) {
if ( retlen ! = je32_to_cpu ( latest_node - > csize ) )
ret = - EIO ;
2005-07-17 11:13:51 +00:00
kfree ( f - > target ) ;
f - > target = NULL ;
2005-03-01 10:50:52 +00:00
up ( & f - > sem ) ;
jffs2_do_clear_inode ( c , f ) ;
return - ret ;
}
2005-07-17 11:13:51 +00:00
f - > target [ je32_to_cpu ( latest_node - > csize ) ] = ' \0 ' ;
2005-09-22 11:25:00 +00:00
dbg_readinode ( " symlink's target '%s' cached \n " , f - > target ) ;
2005-03-01 10:50:52 +00:00
}
2005-11-07 11:16:07 +00:00
2005-04-16 22:20:36 +00:00
/* fall through... */
case S_IFBLK :
case S_IFCHR :
/* Certain inode types should have only one data node, and it's
kept as the metadata node */
if ( f - > metadata ) {
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " Argh. Special inode #%u with mode 0%o had metadata node \n " ,
2005-04-16 22:20:36 +00:00
f - > inocache - > ino , jemode_to_cpu ( latest_node - > mode ) ) ;
up ( & f - > sem ) ;
jffs2_do_clear_inode ( c , f ) ;
return - EIO ;
}
if ( ! frag_first ( & f - > fragtree ) ) {
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " Argh. Special inode #%u with mode 0%o has no fragments \n " ,
2005-04-16 22:20:36 +00:00
f - > inocache - > ino , jemode_to_cpu ( latest_node - > mode ) ) ;
up ( & f - > sem ) ;
jffs2_do_clear_inode ( c , f ) ;
return - EIO ;
}
/* ASSERT: f->fraglist != NULL */
if ( frag_next ( frag_first ( & f - > fragtree ) ) ) {
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " Argh. Special inode #%u with mode 0x%x had more than one node \n " ,
2005-04-16 22:20:36 +00:00
f - > inocache - > ino , jemode_to_cpu ( latest_node - > mode ) ) ;
/* FIXME: Deal with it - check crc32, check for duplicate node, check times and discard the older one */
up ( & f - > sem ) ;
jffs2_do_clear_inode ( c , f ) ;
return - EIO ;
}
/* OK. We're happy */
f - > metadata = frag_first ( & f - > fragtree ) - > node ;
jffs2_free_node_frag ( frag_first ( & f - > fragtree ) ) ;
f - > fragtree = RB_ROOT ;
break ;
}
if ( f - > inocache - > state = = INO_STATE_READING )
jffs2_set_inocache_state ( c , f - > inocache , INO_STATE_PRESENT ) ;
return 0 ;
}
2005-07-27 14:46:14 +00:00
/* Scan the list of all nodes present for this ino, build map of versions, etc. */
2005-11-07 11:16:07 +00:00
int jffs2_do_read_inode ( struct jffs2_sb_info * c , struct jffs2_inode_info * f ,
2005-07-27 14:46:14 +00:00
uint32_t ino , struct jffs2_raw_inode * latest_node )
{
2005-09-22 11:25:00 +00:00
dbg_readinode ( " read inode #%u \n " , ino ) ;
2005-07-27 14:46:14 +00:00
retry_inocache :
spin_lock ( & c - > inocache_lock ) ;
f - > inocache = jffs2_get_ino_cache ( c , ino ) ;
if ( f - > inocache ) {
/* Check its state. We may need to wait before we can use it */
switch ( f - > inocache - > state ) {
case INO_STATE_UNCHECKED :
case INO_STATE_CHECKEDABSENT :
f - > inocache - > state = INO_STATE_READING ;
break ;
2005-11-07 11:16:07 +00:00
2005-07-27 14:46:14 +00:00
case INO_STATE_CHECKING :
case INO_STATE_GC :
/* If it's in either of these states, we need
to wait for whoever ' s got it to finish and
put it back . */
2005-09-22 11:25:00 +00:00
dbg_readinode ( " waiting for ino #%u in state %d \n " , ino , f - > inocache - > state ) ;
2005-07-27 14:46:14 +00:00
sleep_on_spinunlock ( & c - > inocache_wq , & c - > inocache_lock ) ;
goto retry_inocache ;
case INO_STATE_READING :
case INO_STATE_PRESENT :
/* Eep. This should never happen. It can
happen if Linux calls read_inode ( ) again
before clear_inode ( ) has finished though . */
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " Eep. Trying to read_inode #%u when it's already in state %d! \n " , ino , f - > inocache - > state ) ;
2005-07-27 14:46:14 +00:00
/* Fail. That's probably better than allowing it to succeed */
f - > inocache = NULL ;
break ;
default :
BUG ( ) ;
}
}
spin_unlock ( & c - > inocache_lock ) ;
if ( ! f - > inocache & & ino = = 1 ) {
/* Special case - no root inode on medium */
f - > inocache = jffs2_alloc_inode_cache ( ) ;
if ( ! f - > inocache ) {
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " cannot allocate inocache for root inode \n " ) ;
2005-07-27 14:46:14 +00:00
return - ENOMEM ;
}
2005-09-22 11:25:00 +00:00
dbg_readinode ( " creating inocache for root inode \n " ) ;
2005-07-27 14:46:14 +00:00
memset ( f - > inocache , 0 , sizeof ( struct jffs2_inode_cache ) ) ;
f - > inocache - > ino = f - > inocache - > nlink = 1 ;
f - > inocache - > nodes = ( struct jffs2_raw_node_ref * ) f - > inocache ;
f - > inocache - > state = INO_STATE_READING ;
jffs2_add_ino_cache ( c , f - > inocache ) ;
}
if ( ! f - > inocache ) {
2005-07-28 14:46:43 +00:00
JFFS2_ERROR ( " requestied to read an nonexistent ino %u \n " , ino ) ;
2005-07-27 14:46:14 +00:00
return - ENOENT ;
}
return jffs2_do_read_inode_internal ( c , f , latest_node ) ;
}
int jffs2_do_crccheck_inode ( struct jffs2_sb_info * c , struct jffs2_inode_cache * ic )
{
struct jffs2_raw_inode n ;
2006-12-04 23:03:01 +00:00
struct jffs2_inode_info * f = kzalloc ( sizeof ( * f ) , GFP_KERNEL ) ;
2005-07-27 14:46:14 +00:00
int ret ;
if ( ! f )
return - ENOMEM ;
init_MUTEX_LOCKED ( & f - > sem ) ;
f - > inocache = ic ;
ret = jffs2_do_read_inode_internal ( c , f , & n ) ;
if ( ! ret ) {
up ( & f - > sem ) ;
jffs2_do_clear_inode ( c , f ) ;
}
kfree ( f ) ;
return ret ;
}
2005-04-16 22:20:36 +00:00
void jffs2_do_clear_inode ( struct jffs2_sb_info * c , struct jffs2_inode_info * f )
{
struct jffs2_full_dirent * fd , * fds ;
int deleted ;
2006-07-02 14:13:46 +00:00
jffs2_clear_acl ( f ) ;
2006-06-24 00:15:36 +00:00
jffs2_xattr_delete_inode ( c , f - > inocache ) ;
2005-04-16 22:20:36 +00:00
down ( & f - > sem ) ;
deleted = f - > inocache & & ! f - > inocache - > nlink ;
2005-02-27 23:01:36 +00:00
if ( f - > inocache & & f - > inocache - > state ! = INO_STATE_CHECKING )
jffs2_set_inocache_state ( c , f - > inocache , INO_STATE_CLEARING ) ;
2005-04-16 22:20:36 +00:00
if ( f - > metadata ) {
if ( deleted )
jffs2_mark_node_obsolete ( c , f - > metadata - > raw ) ;
jffs2_free_full_dnode ( f - > metadata ) ;
}
jffs2_kill_fragtree ( & f - > fragtree , deleted ? c : NULL ) ;
2005-07-17 11:13:51 +00:00
if ( f - > target ) {
kfree ( f - > target ) ;
f - > target = NULL ;
}
2005-11-07 11:16:07 +00:00
2005-07-17 11:13:51 +00:00
fds = f - > dents ;
while ( fds ) {
fd = fds ;
fds = fd - > next ;
jffs2_free_full_dirent ( fd ) ;
2005-04-16 22:20:36 +00:00
}
2005-02-27 23:01:36 +00:00
if ( f - > inocache & & f - > inocache - > state ! = INO_STATE_CHECKING ) {
2005-04-16 22:20:36 +00:00
jffs2_set_inocache_state ( c , f - > inocache , INO_STATE_CHECKEDABSENT ) ;
2005-02-27 23:01:36 +00:00
if ( f - > inocache - > nodes = = ( void * ) f - > inocache )
jffs2_del_ino_cache ( c , f - > inocache ) ;
}
2005-04-16 22:20:36 +00:00
up ( & f - > sem ) ;
}