rockbox/firmware/common/disk_cache.c

344 lines
10 KiB
C
Raw Permalink Normal View History

Rewrite filesystem code (WIP) This patch redoes the filesystem code from the FAT driver up to the clipboard code in onplay.c. Not every aspect of this is finished therefore it is still "WIP". I don't wish to do too much at once (haha!). What is left to do is get dircache back in the sim and find an implementation for the dircache indicies in the tagcache and playlist code or do something else that has the same benefit. Leaving these out for now does not make anything unusable. All the basics are done. Phone app code should probably get vetted (and app path handling just plain rewritten as environment expansions); the SDL app and Android run well. Main things addressed: 1) Thread safety: There is none right now in the trunk code. Most of what currently works is luck when multiple threads are involved or multiple descriptors to the same file are open. 2) POSIX compliance: Many of the functions behave nothing like their counterparts on a host system. This leads to inconsistent code or very different behavior from native to hosted. One huge offender was rename(). Going point by point would fill a book. 3) Actual running RAM usage: Many targets will use less RAM and less stack space (some more RAM because I upped the number of cache buffers for large memory). There's very little memory lying fallow in rarely-used areas (see 'Key core changes' below). Also, all targets may open the same number of directory streams whereas before those with less than 8MB RAM were limited to 8, not 12 implying those targets will save slightly less. 4) Performance: The test_disk plugin shows markedly improved performance, particularly in the area of (uncached) directory scanning, due partly to more optimal directory reading and to a better sector cache algorithm. Uncached times tend to be better while there is a bit of a slowdown in dircache due to it being a bit heavier of an implementation. It's not noticeable by a human as far as I can say. Key core changes: 1) Files and directories share core code and data structures. 2) The filesystem code knows which descriptors refer to same file. This ensures that changes from one stream are appropriately reflected in every open descriptor for that file (fileobj_mgr.c). 3) File and directory cache buffers are borrowed from the main sector cache. This means that when they are not in use by a file, they are not wasted, but used for the cache. Most of the time, only a few of them are needed. It also means that adding more file and directory handles is less expensive. All one must do in ensure a large enough cache to borrow from. 4) Relative path components are supported and the namespace is unified. It does not support full relative paths to an implied current directory; what is does support is use of "." and "..". Adding the former would not be very difficult. The namespace is unified in the sense that volumes may be specified several times along with relative parts, e.g.: "/<0>/foo/../../<1>/bar" :<=> "/<1>/bar". 5) Stack usage is down due to sharing of data, static allocation and less duplication of strings on the stack. This requires more serialization than I would like but since the number of threads is limited to a low number, the tradoff in favor of the stack seems reasonable. 6) Separates and heirarchicalizes (sic) the SIM and APP filesystem code. SIM path and volume handling is just like the target. Some aspects of the APP file code get more straightforward (e.g. no path hashing is needed). Dircache: Deserves its own section. Dircache is new but pays homage to the old. The old one was not compatible and so it, since it got redone, does all the stuff it always should have done such as: 1) It may be update and used at any time during the build process. No longer has one to wait for it to finish building to do basic file management (create, remove, rename, etc.). 2) It does not need to be either fully scanned or completely disabled; it can be incomplete (i.e. overfilled, missing paths), still be of benefit and be correct. 3) Handles mounting and dismounting of individual volumes which means a full rebuild is not needed just because you pop a new SD card in the slot. Now, because it reuses its freed entry data, may rebuild only that volume. 4) Much more fundamental to the file code. When it is built, it is the keeper of the master file list whether enabled or not ("disabled" is just a state of the cache). Its must always to ready to be started and bind all streams opened prior to being enabled. 5) Maintains any short filenames in OEM format which means that it does not need to be rebuilt when changing the default codepage. Miscellaneous Compatibility: 1) Update any other code that would otherwise not work such as the hotswap mounting code in various card drivers. 2) File management: Clipboard needed updating because of the behavioral changes. Still needs a little more work on some finer points. 3) Remove now-obsolete functionality such as the mutex's "no preempt" flag (which was only for the prior FAT driver). 4) struct dirinfo uses time_t rather than raw FAT directory entry time fields. I plan to follow up on genericizing everything there (i.e. no FAT attributes). 5) unicode.c needed some redoing so that the file code does not try try to load codepages during a scan, which is actually a problem with the current code. The default codepage, if any is required, is now kept in RAM separarately (bufalloced) from codepages specified to iso_decode() (which must not be bufalloced because the conversion may be done by playback threads). Brings with it some additional reusable core code: 1) Revised file functions: Reusable code that does things such as safe path concatenation and parsing without buffer limitations or data duplication. Variants that copy or alter the input path may be based off these. To do: 1) Put dircache functionality back in the sim. Treating it internally as a different kind of file system seems the best approach at this time. 2) Restore use of dircache indexes in the playlist and database or something effectively the same. Since the cache doesn't have to be complete in order to be used, not getting a hit on the cache doesn't unambiguously say if the path exists or not. Change-Id: Ia30f3082a136253e3a0eae0784e3091d138915c8 Reviewed-on: http://gerrit.rockbox.org/566 Reviewed-by: Michael Sevakis <jethead71@rockbox.org> Tested: Michael Sevakis <jethead71@rockbox.org>
2013-08-06 02:02:45 +00:00
/***************************************************************************
* __________ __ ___.
* Open \______ \ ____ ____ | | _\_ |__ _______ ___
* Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ /
* Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < <
* Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \
* \/ \/ \/ \/ \/
* $Id$
*
* Copyright (C) 2014 by Michael Sevakis
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version 2
* of the License, or (at your option) any later version.
*
* This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY
* KIND, either express or implied.
*
****************************************************************************/
#include "config.h"
#include "debug.h"
#include "system.h"
#include "linked_list.h"
#include "disk_cache.h"
#include "fs_defines.h"
Rewrite filesystem code (WIP) This patch redoes the filesystem code from the FAT driver up to the clipboard code in onplay.c. Not every aspect of this is finished therefore it is still "WIP". I don't wish to do too much at once (haha!). What is left to do is get dircache back in the sim and find an implementation for the dircache indicies in the tagcache and playlist code or do something else that has the same benefit. Leaving these out for now does not make anything unusable. All the basics are done. Phone app code should probably get vetted (and app path handling just plain rewritten as environment expansions); the SDL app and Android run well. Main things addressed: 1) Thread safety: There is none right now in the trunk code. Most of what currently works is luck when multiple threads are involved or multiple descriptors to the same file are open. 2) POSIX compliance: Many of the functions behave nothing like their counterparts on a host system. This leads to inconsistent code or very different behavior from native to hosted. One huge offender was rename(). Going point by point would fill a book. 3) Actual running RAM usage: Many targets will use less RAM and less stack space (some more RAM because I upped the number of cache buffers for large memory). There's very little memory lying fallow in rarely-used areas (see 'Key core changes' below). Also, all targets may open the same number of directory streams whereas before those with less than 8MB RAM were limited to 8, not 12 implying those targets will save slightly less. 4) Performance: The test_disk plugin shows markedly improved performance, particularly in the area of (uncached) directory scanning, due partly to more optimal directory reading and to a better sector cache algorithm. Uncached times tend to be better while there is a bit of a slowdown in dircache due to it being a bit heavier of an implementation. It's not noticeable by a human as far as I can say. Key core changes: 1) Files and directories share core code and data structures. 2) The filesystem code knows which descriptors refer to same file. This ensures that changes from one stream are appropriately reflected in every open descriptor for that file (fileobj_mgr.c). 3) File and directory cache buffers are borrowed from the main sector cache. This means that when they are not in use by a file, they are not wasted, but used for the cache. Most of the time, only a few of them are needed. It also means that adding more file and directory handles is less expensive. All one must do in ensure a large enough cache to borrow from. 4) Relative path components are supported and the namespace is unified. It does not support full relative paths to an implied current directory; what is does support is use of "." and "..". Adding the former would not be very difficult. The namespace is unified in the sense that volumes may be specified several times along with relative parts, e.g.: "/<0>/foo/../../<1>/bar" :<=> "/<1>/bar". 5) Stack usage is down due to sharing of data, static allocation and less duplication of strings on the stack. This requires more serialization than I would like but since the number of threads is limited to a low number, the tradoff in favor of the stack seems reasonable. 6) Separates and heirarchicalizes (sic) the SIM and APP filesystem code. SIM path and volume handling is just like the target. Some aspects of the APP file code get more straightforward (e.g. no path hashing is needed). Dircache: Deserves its own section. Dircache is new but pays homage to the old. The old one was not compatible and so it, since it got redone, does all the stuff it always should have done such as: 1) It may be update and used at any time during the build process. No longer has one to wait for it to finish building to do basic file management (create, remove, rename, etc.). 2) It does not need to be either fully scanned or completely disabled; it can be incomplete (i.e. overfilled, missing paths), still be of benefit and be correct. 3) Handles mounting and dismounting of individual volumes which means a full rebuild is not needed just because you pop a new SD card in the slot. Now, because it reuses its freed entry data, may rebuild only that volume. 4) Much more fundamental to the file code. When it is built, it is the keeper of the master file list whether enabled or not ("disabled" is just a state of the cache). Its must always to ready to be started and bind all streams opened prior to being enabled. 5) Maintains any short filenames in OEM format which means that it does not need to be rebuilt when changing the default codepage. Miscellaneous Compatibility: 1) Update any other code that would otherwise not work such as the hotswap mounting code in various card drivers. 2) File management: Clipboard needed updating because of the behavioral changes. Still needs a little more work on some finer points. 3) Remove now-obsolete functionality such as the mutex's "no preempt" flag (which was only for the prior FAT driver). 4) struct dirinfo uses time_t rather than raw FAT directory entry time fields. I plan to follow up on genericizing everything there (i.e. no FAT attributes). 5) unicode.c needed some redoing so that the file code does not try try to load codepages during a scan, which is actually a problem with the current code. The default codepage, if any is required, is now kept in RAM separarately (bufalloced) from codepages specified to iso_decode() (which must not be bufalloced because the conversion may be done by playback threads). Brings with it some additional reusable core code: 1) Revised file functions: Reusable code that does things such as safe path concatenation and parsing without buffer limitations or data duplication. Variants that copy or alter the input path may be based off these. To do: 1) Put dircache functionality back in the sim. Treating it internally as a different kind of file system seems the best approach at this time. 2) Restore use of dircache indexes in the playlist and database or something effectively the same. Since the cache doesn't have to be complete in order to be used, not getting a hit on the cache doesn't unambiguously say if the path exists or not. Change-Id: Ia30f3082a136253e3a0eae0784e3091d138915c8 Reviewed-on: http://gerrit.rockbox.org/566 Reviewed-by: Michael Sevakis <jethead71@rockbox.org> Tested: Michael Sevakis <jethead71@rockbox.org>
2013-08-06 02:02:45 +00:00
#include "bitarray.h"
/* Cache: LRU cache with separately-chained hashtable
*
* Each entry of the map is the mapped location of the hashed sector value
* where each bit in each map entry indicates which corresponding cache
* entries are occupied by sector values that collide in that map entry.
*
* Each volume is given its own bit map.
*
* To probe for a specific key, each bit in the map entry must be examined,
* its position used as an index into the cache_entry array and the actual
* sector information compared for that cache entry. If the search exhausts
* all bits, the sector is not cached.
*
* To avoid long chains, the map entry count should be much greater than the
* number of cache entries. Since the cache is an LRU design, no buffer entry
* in the array is intrinsically associated with any particular sector number
* or volume.
*
* Example 6-sector cache with 8-entry map:
* cache entry 543210
* cache map 100000 <- sector number hashes into map
* 000000
* 000100
* 000000
* 010000
* 000000
* 001001 <- collision
* 000000
* volume map 111101 <- entry usage by the volume (OR of all map entries)
*/
enum dce_flags /* flags for each cache entry */
{
DCE_INUSE = 0x01, /* entry in use and valid */
DCE_DIRTY = 0x02, /* entry is dirty in need of writeback */
DCE_BUF = 0x04, /* entry is being used as a general buffer */
};
struct disk_cache_entry
{
struct lldc_node node; /* LRU list links */
unsigned char flags; /* entry flags */
#ifdef HAVE_MULTIVOLUME
unsigned char volume; /* volume of sector */
#endif
unsigned long sector; /* cached disk sector number */
};
BITARRAY_TYPE_DECLARE(cache_map_entry_t, cache_map, DC_NUM_ENTRIES)
static inline unsigned int map_sector(unsigned long sector)
{
/* keep sector hash simple for now */
return sector % DC_MAP_NUM_ENTRIES;
}
static struct lldc_head cache_lru; /* LRU cache list (head = LRU item) */
static struct disk_cache_entry cache_entry[DC_NUM_ENTRIES];
static cache_map_entry_t cache_map_entry[NUM_VOLUMES][DC_MAP_NUM_ENTRIES];
static cache_map_entry_t cache_vol_map[NUM_VOLUMES] IBSS_ATTR;
static uint8_t cache_buffer[DC_NUM_ENTRIES][DC_CACHE_BUFSIZE] CACHEALIGN_ATTR;
struct mutex disk_cache_mutex SHAREDBSS_ATTR;
#define CACHE_MAP_ENTRY(volume, mapnum) \
cache_map_entry[IF_MV_VOL(volume)][mapnum]
#define CACHE_VOL_MAP(volume) \
cache_vol_map[IF_MV_VOL(volume)]
#define DCE_LRU() ((struct disk_cache_entry *)cache_lru.head)
#define DCE_NEXT(fce) ((struct disk_cache_entry *)(fce)->node.next)
#define NODE_DCE(node) ((struct disk_cache_entry *)(node))
/* get the cache index from a pointer to a buffer */
#define DCIDX_FROM_BUF(buf) \
((uint8_t (*)[DC_CACHE_BUFSIZE])(buf) - cache_buffer)
#define DCIDX_FROM_DCE(dce) \
((dce) - cache_entry)
/* set the in-use bit in the map */
static inline void cache_bitmap_set_bit(int volume, unsigned int mapnum,
unsigned int bitnum)
{
cache_map_set_bit(&CACHE_MAP_ENTRY(volume, mapnum), bitnum);
cache_map_set_bit(&CACHE_VOL_MAP(volume), bitnum);
(void)volume;
}
/* clear the in-use bit in the map */
static inline void cache_bitmap_clear_bit(int volume, unsigned int mapnum,
unsigned int bitnum)
{
cache_map_clear_bit(&CACHE_MAP_ENTRY(volume, mapnum), bitnum);
cache_map_clear_bit(&CACHE_VOL_MAP(volume), bitnum);
(void)volume;
}
/* make entry MRU by moving it to the list tail */
static inline void touch_cache_entry(struct disk_cache_entry *which)
{
struct lldc_node *lru = cache_lru.head;
struct lldc_node *node = &which->node;
if (node == lru->prev) /* already MRU */
; /**/
else if (node == lru) /* is the LRU? just rotate list */
cache_lru.head = lru->next;
else /* somewhere else; move it */
{
lldc_remove(&cache_lru, node);
lldc_insert_last(&cache_lru, node);
}
}
/* remove LRU entry from the cache list to use as a buffer */
static struct disk_cache_entry * cache_remove_lru_entry(void)
{
struct lldc_node *lru = cache_lru.head;
/* at least one is reserved for client */
if (lru == lru->next)
return NULL;
/* remove it; next-LRU becomes the LRU */
lldc_remove(&cache_lru, lru);
return NODE_DCE(lru);
}
/* return entry to the cache list and set it LRU */
static void cache_return_lru_entry(struct disk_cache_entry *fce)
{
lldc_insert_first(&cache_lru, &fce->node);
}
/* discard the entry's data and mark it unused */
static inline void cache_discard_entry(struct disk_cache_entry *dce,
unsigned int index)
{
cache_bitmap_clear_bit(IF_MV_VOL(dce->volume), map_sector(dce->sector),
index);
dce->flags = 0;
}
/* search the cache for the specified sector, returning a buffer, either
to the specified sector, if it exists, or a new/evicted entry that must
be filled */
void * dc_cache_probe(IF_MV(int volume,) unsigned long sector,
unsigned int *flagsp)
{
unsigned int mapnum = map_sector(sector);
FOR_EACH_BITARRAY_SET_BIT(&CACHE_MAP_ENTRY(volume, mapnum), index)
{
struct disk_cache_entry *dce = &cache_entry[index];
if (dce->sector == sector)
{
*flagsp = DCE_INUSE;
touch_cache_entry(dce);
return cache_buffer[index];
}
}
/* sector not found so the LRU is the victim */
struct disk_cache_entry *dce = DCE_LRU();
cache_lru.head = dce->node.next;
unsigned int index = DCIDX_FROM_DCE(dce);
void *buf = cache_buffer[index];
unsigned int old_flags = dce->flags;
if (old_flags)
{
int old_volume = IF_MV_VOL(dce->volume);
unsigned long sector = dce->sector;
unsigned int old_mapnum = map_sector(sector);
if (old_flags & DCE_DIRTY)
dc_writeback_callback(IF_MV(old_volume,) sector, buf);
if (mapnum == old_mapnum IF_MV( && volume == old_volume ))
goto finish_setup;
cache_bitmap_clear_bit(old_volume, old_mapnum, index);
}
cache_bitmap_set_bit(IF_MV_VOL(volume), mapnum, index);
finish_setup:
dce->flags = DCE_INUSE;
#ifdef HAVE_MULTIVOLUME
dce->volume = volume;
#endif
dce->sector = sector;
*flagsp = 0;
return buf;
}
/* mark in-use cache entry as dirty by buffer */
void dc_dirty_buf(void *buf)
{
unsigned int index = DCIDX_FROM_BUF(buf);
if (index >= DC_NUM_ENTRIES)
return;
/* dirt remains, sticky until flushed */
struct disk_cache_entry *fce = &cache_entry[index];
if (fce->flags & DCE_INUSE)
fce->flags |= DCE_DIRTY;
}
/* discard in-use cache entry by buffer */
void dc_discard_buf(void *buf)
{
unsigned int index = DCIDX_FROM_BUF(buf);
if (index >= DC_NUM_ENTRIES)
return;
struct disk_cache_entry *dce = &cache_entry[index];
if (dce->flags & DCE_INUSE)
cache_discard_entry(dce, index);
}
/* commit all dirty cache entries to storage for a specified volume */
void dc_commit_all(IF_MV_NONVOID(int volume))
{
DEBUGF("dc_commit_all()\n");
FOR_EACH_BITARRAY_SET_BIT(&CACHE_VOL_MAP(volume), index)
{
struct disk_cache_entry *dce = &cache_entry[index];
unsigned int flags = dce->flags;
if (flags & DCE_DIRTY)
{
dc_writeback_callback(IF_MV(volume,) dce->sector,
cache_buffer[index]);
dce->flags = flags & ~DCE_DIRTY;
}
}
}
/* discard all cache entries from the specified volume */
void dc_discard_all(IF_MV_NONVOID(int volume))
{
DEBUGF("dc_discard_all()\n");
FOR_EACH_BITARRAY_SET_BIT(&CACHE_VOL_MAP(volume), index)
cache_discard_entry(&cache_entry[index], index);
}
/* expropriate a buffer from the cache */
void * dc_get_buffer(void)
{
dc_lock_cache();
void *buf = NULL;
struct disk_cache_entry *dce = cache_remove_lru_entry();
if (dce)
{
unsigned int index = DCIDX_FROM_DCE(dce);
unsigned int flags = dce->flags;
buf = cache_buffer[index];
if (flags)
{
/* must first commit this sector if dirty */
if (flags & DCE_DIRTY)
dc_writeback_callback(IF_MV(dce->volume,) dce->sector, buf);
cache_discard_entry(dce, index);
}
dce->flags = DCE_BUF;
}
/* cache is out of buffers */
dc_unlock_cache();
return buf;
}
/* return buffer to the cache by buffer */
void dc_release_buffer(void *buf)
{
unsigned int index = DCIDX_FROM_BUF(buf);
if (index >= DC_NUM_ENTRIES)
return;
dc_lock_cache();
struct disk_cache_entry *dce = &cache_entry[index];
if (dce->flags & DCE_BUF)
{
dce->flags = 0;
cache_return_lru_entry(dce);
}
dc_unlock_cache();
}
/* one-time init at startup */
void dc_init(void)
{
mutex_init(&disk_cache_mutex);
lldc_init(&cache_lru);
for (unsigned int i = 0; i < DC_NUM_ENTRIES; i++)
lldc_insert_last(&cache_lru, &cache_entry[i].node);
}