kolibrios/kernel/branches/Kolibri-acpi/blkdev/disk.inc

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;                                                              ;;
;; Copyright (C) KolibriOS team 2011. All rights reserved.      ;;
;; Distributed under terms of the GNU General Public License    ;;
;;                                                              ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

$Revision: 2257 $

; =============================================================================
; ================================= Constants =================================
; =============================================================================
; Error codes for callback functions.
DISK_STATUS_OK		    = 0 ; success
DISK_STATUS_GENERAL_ERROR   = -1; if no other code is suitable
DISK_STATUS_INVALID_CALL    = 1 ; invalid input parameters
DISK_STATUS_NO_MEDIA	    = 2 ; no media present
DISK_STATUS_END_OF_MEDIA    = 3 ; end of media while reading/writing data
; Driver flags. Represent bits in DISK.DriverFlags.
DISK_NO_INSERT_NOTIFICATION = 1
; Media flags. Represent bits in DISKMEDIAINFO.Flags.
DISK_MEDIA_READONLY = 1

; If too many partitions are detected,there is probably an error on the disk.
; 256 partitions should be enough for any reasonable use.
; Also, the same number is limiting the number of MBRs to process; if
; too many MBRs are visible,there probably is a loop in the MBR structure.
MAX_NUM_PARTITIONS = 256

; =============================================================================
; ================================ Structures =================================
; =============================================================================
; This structure defines all callback functions for working with the physical
; device. They are implemented by a driver. Objects with this structure reside
; in a driver.
struct DISKFUNC
.strucsize	dd	?
; Size of the structure. This field is intended for possible extensions of
; this structure. If a new function is added to this structure and a driver
; implements an old version, the caller can detect this by checking .strucsize,
; so the driver remains compatible.
.close		dd	?
; The pointer to the function which frees all driver-specific resources for
; the disk.
; Optional, may be NULL.
; void close(void* userdata);
.closemedia	dd	?
; The pointer to the function which informs the driver that the kernel has
; finished all processing with the current media. If media is removed, the
; driver should decline all requests to that media with DISK_STATUS_NO_MEDIA,
; even if new media is inserted, until this function is called. If media is
; removed, a new call to 'disk_media_changed' is not allowed until this
; function is called.
; Optional, may be NULL (if media is not removable).
; void closemedia(void* userdata);
.querymedia	dd	?
; The pointer to the function which determines capabilities of the media.
; int querymedia(void* userdata, DISKMEDIAINFO* info);
; Return value: one of DISK_STATUS_*
.read		dd	?
; The pointer to the function which reads data from the device.
; int read(void* userdata, void* buffer, __int64 startsector, int* numsectors);
; input: *numsectors = number of sectors to read
; output: *numsectors = number of sectors which were successfully read
; Return value: one of DISK_STATUS_*
.write		dd	?
; The pointer to the function which writes data to the device.
; Optional, may be NULL.
; int write(void* userdata, void* buffer, __int64 startsector, int* numsectors);
; input: *numsectors = number of sectors to write
; output: *numsectors = number of sectors which were successfully written
; Return value: one of DISK_STATUS_*
.flush		dd	?
; The pointer to the function which flushes the internal device cache.
; Optional, may be NULL.
; int flush(void* userdata);
; Return value: one of DISK_STATUS_*
; Note that read/write are called by the cache manager, so a driver should not
; create a software cache. This function is implemented for flushing a hardware
; cache, if it exists.
.adjust_cache_size	dd	?
; The pointer to the function which returns the cache size for this device.
; Optional, may be NULL.
; unsigned int adjust_cache_size(unsigned int suggested_size);
; Return value: 0 = disable cache, otherwise = used cache size in bytes.
ends

; This structure holds information on a medium.
; Objects with this structure are allocated by the kernel as a part of the DISK
; structure and are filled by a driver in the 'querymedia' callback.
struct DISKMEDIAINFO
.Flags		dd	?
; Combination of DISK_MEDIA_* bits.
.SectorSize	dd	?
; Size of the sector.
.Capacity	dq	?
; Size of the media in sectors.
ends

; This structure represents the disk cache. To follow the old implementation,
; there are two distinct caches for a disk, one for "system" data,and the other
; for "application" data.
struct DISKCACHE
.Lock		MUTEX
; Lock to protect the cache.
; The following fields are inherited from data32.inc:cache_ideX.
.pointer     rd 1
.data_size   rd 1   ; not use
.data	     rd 1
.sad_size    rd 1
.search_start	    rd 1
ends

; This structure represents a disk device and its media for the kernel.
; This structure is allocated by the kernel in the 'disk_add' function,
; freed in the 'disk_dereference' function.
struct DISK
; Fields of disk object
.Next		dd	?
.Prev		dd	?
; All disk devices are linked in one list with these two fields.
; Head of the list is the 'disk_list' variable.
.Functions	dd	?
; Pointer to the 'DISKFUNC' structure with driver functions.
.Name		dd	?
; Pointer to the string used for accesses through the global filesystem.
.UserData	dd	?
; This field is passed to all callback functions so a driver can decide which
; physical device is addressed.
.DriverFlags	dd	?
; Bitfield. Currently only DISK_NO_INSERT_NOTIFICATION bit is defined.
; If it is set, the driver will never issue 'disk_media_changed' notification
; with argument set to true, so the kernel must try to detect media during
; requests from the file system.
.RefCount	dd	?
; Count of active references to this structure. One reference is kept during
; the lifetime of the structure between 'disk_add' and 'disk_del'.
; Another reference is taken during any filesystem operation for this disk.
; One reference is added if media is inserted.
; The structure is destroyed when the reference count decrements to zero:
; this usually occurs in 'disk_del', but can be delayed to the end of last
; filesystem operation, if one is active.
.MediaLock		MUTEX
; Lock to protect the MEDIA structure. See the description after
; 'disk_list_mutex' for the locking strategy.
; Fields of media object
.MediaInserted		db	?
; 0 if media is not inserted, nonzero otherwise.
.MediaUsed		db	?
; 0 if media fields are not used, nonzero otherwise. If .MediaRefCount is
; nonzero, this field is nonzero too; however, when .MediaRefCount goes
; to zero, there is some time interval during which media object is still used.
		align 4
; The following fields are not valid unless either .MediaInserted is nonzero
; or they are accessed from a code which has obtained the reference when
; .MediaInserted was nonzero.
.MediaRefCount		dd	?
; Count of active references to the media object. One reference is kept during
; the lifetime of the media between two calls to 'disk_media_changed'.
; Another reference is taken during any filesystem operation for this media.
; The callback 'closemedia' is called when the reference count decrements to
; zero: this usually occurs in 'disk_media_changed', but can be delayed to the
; end of the last filesystem operation, if one is active.
.MediaInfo		DISKMEDIAINFO
; This field keeps information on the current media.
.NumPartitions	dd	?
; Number of partitions on this media.
.Partitions	dd	?
; Pointer to array of .NumPartitions pointers to PARTITION structures.
.cache_size	dd	?
; inherited from cache_ideX_size
.SysCache	DISKCACHE
.AppCache	DISKCACHE
; Two caches for the disk.
ends

; This structure represents one partition for the kernel. This is a base
; template, the actual contents after common fields is determined by the
; file system code for this partition.
struct PARTITION
.FirstSector	dq	?
; First sector of the partition.
.Length 	dq	?
; Length of the partition in sectors.
.Disk		dd	?
; Pointer to parent DISK structure.
.FSUserFunctions	dd	?
; Handlers for the sysfunction 70h. This field is a pointer to the following
; array. The first dword is a number of supported subfunctions, other dwords
; point to handlers of corresponding subfunctions.
; This field is 0 if file system is not recognized.
; ...fs-specific data may follow...
ends

; This is an external structure, it represents an entry in the partition table.
struct PARTITION_TABLE_ENTRY
.Bootable	db	?
; 80h = bootable partition, 0 = non-bootable partition, other values = invalid
.FirstHead	db	?
.FirstSector	db	?
.FirstTrack	db	?
; Coordinates of first sector in CHS.
.Type		db	?
; Partition type, one of predefined constants. 0 = empty, several types denote
; extended partition (see process_partition_table_entry), we are not interested
; in other values.
.LastHead	db	?
.LastSector	db	?
.LastTrack	db	?
; Coordinates of last sector in CHS.
.FirstAbsSector dd	?
; Coordinate of first sector in LBA.
.Length 	dd	?
; Length of the partition in sectors.
ends

; =============================================================================
; ================================ Global data ================================
; =============================================================================
iglobal
; The pseudo-item for the list of all DISK structures.
; Initialized to the empty list.
disk_list:
	dd	disk_list
	dd	disk_list
endg
uglobal
; This mutex guards all operations with the global list of DISK structures.
disk_list_mutex MUTEX
; * There are two dependent objects, a disk and a media. In the simplest case,
;   disk and media are both non-removable. However, in the general case both
;   can be removed at any time, simultaneously or only media,and this makes things
;   complicated.
; * For efficiency, both disk and media objects are located in the one
;   structure named DISK. However, logically they are different.
; * The following operations use data of disk object: adding (disk_add);
;   deleting (disk_del); filesystem (fs_lfn which eventually calls
;   dyndisk_handler or dyndisk_enum_root).
; * The following operations use data of media object: adding/removing
;   (disk_media_changed); filesystem (fs_lfn which eventually calls
;   dyndisk_handler; dyndisk_enum_root doesn't work with media).
; * Notifications disk_add, disk_media_changed, disk_del are synchronized
;   between themselves, this is a requirement for the driver. However, file
;   system operations are asynchronous, can be issued at any time by any
;   thread.
; * We must prevent a situation when a filesystem operation thinks that the
;   object is still valid but in fact the notification has destroyed the
;   object. So we keep a reference counter for both disk and media and destroy
;   the object when this counter goes to zero.
; * The driver must know when it is safe to free driver-allocated resources.
;   The object can be alive even after death notification has completed.
;   We use special callbacks to satisfy both assertions: 'close' for the disk
;   and 'closemedia' for the media. The destruction of the object includes
;   calling the corresponding callback.
; * Each filesystem operation keeps one reference for the disk and one
;   reference for the media. Notification disk_del forces notification on the
;   media death, so the reference counter for the disk is always not less than
;   the reference counter for the media.
; * Two operations "get the object" and "increment the reference counter" can
;   not be done simultaneously. We use a mutex to guard the consistency here.
;   It must be a part of the container for the object, so that this mutex can
;   be acquired as a part of getting the object from the container. The
;   container for disk object is the global list, and this list is guarded by
;   'disk_list_mutex'. The container for media object is the disk object, and
;   the corresponding mutex is DISK.MediaLock.
; * Notifications do not change the data of objects, they can only remove
;   objects. Thus we don't need another synchronization at this level. If two
;   filesystem operations are referencing the same filesystem data, this is
;   better resolved at the level of the filesystem.
endg

iglobal
; The function 'disk_scan_partitions' needs three 512-byte buffers for
; MBR, bootsector and fs-temporary sector data. It can not use the static
; buffers always, since it can be called for two or more disks in parallel.
; However, this case is not typical. We reserve three static 512-byte buffers
; and a flag that these buffers are currently used. If 'disk_scan_partitions'
; detects that the buffers are currently used, it allocates buffers from the
; heap.
; The flag is implemented as a global dword variable. When the static buffers
; are not used, the value is -1. When the static buffers are used, the value
; is normally 0 and temporarily can become greater. The function increments
; this value. If the resulting value is zero, it uses the buffers and
; decrements the value when the job is done. Otherwise, it immediately
; decrements the value and uses buffers from the heap, allocated in the
; beginning and freed in the end.
partition_buffer_users	dd	-1
endg
uglobal
; The static buffers for MBR, bootsector and fs-temporary sector data.
align 16
mbr_buffer	rb	512
bootsect_buffer rb	512
fs_tmp_buffer	rb	512
endg

iglobal
; This is the array of default implementations of driver callbacks.
; Same as DRIVERFUNC structure except for the first field; all functions must
; have the default implementations.
align 4
disk_default_callbacks:
	dd	disk_default_close
	dd	disk_default_closemedia
	dd	disk_default_querymedia
	dd	disk_default_read
	dd	disk_default_write
	dd	disk_default_flush
	dd	disk_default_adjust_cache_size
endg

; =============================================================================
; ================================= Functions =================================
; =============================================================================

; This function registers a disk device.
; This includes:
; - allocating an internal structure describing this device;
; - registering this structure in the global filesystem.
; The function initializes the disk as if there is no media. If a media is
; present, the function 'disk_media_changed' should be called after this
; function succeeds.
; Parameters:
; [esp+4] = pointer to DISKFUNC structure with the callbacks
; [esp+8] = pointer to name (ASCIIZ string)
; [esp+12] = userdata to be passed to the callbacks as is.
; [esp+16] = flags, bitfield. Currently only DISK_NO_INSERT_NOTIFICATION bit
;            is defined.
; Return value:
; NULL = operation has failed
; non-NULL = handle of the disk. This handle can be used
; in the operations with other Disk* functions.
; The handle is the pointer to the internal structure DISK.
disk_add:
	push	ebx esi 	; save used registers to be stdcall
; 1. Allocate the DISK structure.
; 1a. Call the heap manager.
	push	sizeof.DISK
	pop	eax
	call	malloc
; 1b. Check the result. If allocation failed, return (go to 9) with eax = 0.
	test	eax, eax
	jz	.nothing
; 2. Copy the disk name to the DISK structure.
; 2a. Get length of the name, including the terminating zero.
	mov	ebx, [esp+8+8]	; ebx = pointer to name
	push	eax		; save allocated pointer to DISK
	xor	eax, eax	; the argument of malloc() is in eax
@@:
	inc	eax
	cmp	byte [ebx+eax-1], 0
	jnz	@b
; 2b. Call the heap manager.
	call	malloc
; 2c. Check the result. If allocation failed, go to 7.
	pop	esi		; restore allocated pointer to DISK
	test	eax, eax
	jz	.free
; 2d. Store the allocated pointer to the DISK structure.
	mov	[esi+DISK.Name], eax
; 2e. Copy the name.
@@:
	mov	dl, [ebx]
	mov	[eax], dl
	inc	ebx
	inc	eax
	test	dl, dl
	jnz	@b
; 3. Copy other arguments of the function to the DISK structure.
	mov	eax, [esp+4+8]
	mov	[esi+DISK.Functions], eax
	mov	eax, [esp+12+8]
	mov	[esi+DISK.UserData], eax
	mov	eax, [esp+16+8]
	mov	[esi+DISK.DriverFlags], eax
; 4. Initialize other fields of the DISK structure.
; Media is not inserted, reference counter is 1.
	lea	ecx, [esi+DISK.MediaLock]
	call	mutex_init
	xor	eax, eax
	mov	dword [esi+DISK.MediaInserted], eax
	inc	eax
	mov	[esi+DISK.RefCount], eax
; The DISK structure is initialized.
; 5. Insert the new structure to the global list.
; 5a. Acquire the mutex.
	mov	ecx, disk_list_mutex
	call	mutex_lock
; 5b. Insert item to the tail of double-linked list.
	mov	edx, disk_list
	list_add_tail esi, edx	   ;esi= new edx= list head
; 5c. Release the mutex.
	call	mutex_unlock
; 6. Return with eax = pointer to DISK.
	xchg	eax, esi
	jmp	.nothing
.free:
; Memory allocation for DISK structure succeeded, but for disk name failed.
; 7. Free the DISK structure.
	xchg	eax, esi
	call	free
; 8. Return with eax = 0.
	xor	eax, eax
.nothing:
; 9. Return.
	pop	esi ebx 	; restore used registers to be stdcall
	ret	16		; purge 4 dword arguments to be stdcall

; This function deletes a disk device from the global filesystem.
; This includes:
; - removing a media including all partitions;
; - deleting this structure from the global filesystem;
; - dereferencing the DISK structure and possibly destroying it.
; Parameters:
; [esp+4] = handle of the disk, i.e. the pointer to the DISK structure.
; Return value: none.
disk_del:
	push	esi	    ; save used registers to be stdcall
; 1. Force media to be removed. If the media is already removed, the
; call does nothing.
	mov	esi, [esp+4+8]	; esi = handle of the disk
	stdcall disk_media_changed, esi, 0
; 2. Delete the structure from the global list.
; 2a. Acquire the mutex.
	mov	ecx, disk_list_mutex
	call	mutex_lock
; 2b. Delete item from double-linked list.
	mov	eax, [esi+DISK.Next]
	mov	edx, [esi+DISK.Prev]
	mov	[eax+DISK.Prev], edx
	mov	[edx+DISK.Next], eax
; 2c. Release the mutex.
	call	mutex_unlock
; 3. The structure still has one reference created in disk_add. Remove this
; reference. If there are no other references, disk_dereference will free the
; structure.
	call	disk_dereference
; 4. Return.
	pop	esi		; restore used registers to be stdcall
	ret	4		; purge 1 dword argument to be stdcall

; This is an internal function which removes a previously obtained reference
; to the disk. If this is the last reference, this function lets the driver
; finalize all associated data, and afterwards frees the DISK structure.
; esi = pointer to DISK structure
disk_dereference:
; 1. Decrement reference counter. Use atomic operation to correctly handle
; possible simultaneous calls.
lock	dec	[esi+DISK.RefCount]
; 2. If the result is nonzero, there are other references, so nothing to do.
; In this case, return (go to 4).
	jnz	.nothing
; 3. If we are here, we just removed the last reference and must destroy the
; disk object.
; 3a. Call the driver.
	mov	al, DISKFUNC.close
	stdcall disk_call_driver
; 3b. Free the structure.
	xchg	eax, esi
	call	free
; 4. Return.
.nothing:
	ret

; This is an internal function which removes a previously obtained reference
; to the media. If this is the last reference, this function calls 'closemedia'
; callback to signal the driver that the processing has finished and it is safe
; to inform about a new media.
; esi = pointer to DISK structure
disk_media_dereference:
; 1. Decrement reference counter. Use atomic operation to correctly handle
; possible simultaneous calls.
lock	dec	[esi+DISK.MediaRefCount]
; 2. If the result is nonzero, there are other references, so nothing to do.
; In this case, return (go to 4).
	jnz	.nothing
; 3. If we are here, we just removed the last reference and must destroy the
; media object.
; Note that the same place inside the DISK structure is reused for all media
; objects, so we must guarantee that reusing does not happen while freeing.
; Reusing is only possible when someone processes a new media. There are two
; mutually exclusive variants:
; * driver issues media insert notifications (DISK_NO_INSERT_NOTIFICATION bit
;   in DISK.DriverFlags is not set). In this case, we require from the driver
;   that such notification (except for the first one) can occur only after a
;   call to 'closemedia' callback.
; * driver does not issue media insert notifications. In this case, the kernel
;   itself must sometimes check whether media is inserted. We have the flag
;   DISK.MediaUsed, visible to the kernel. This flag signals to the other parts
;   of kernel that the way is free.
; In the first case other parts of the kernel do not use DISK.MediaUsed, so it
; does not matter when this flag is cleared. In the second case this flag must
; be cleared after all other actions, including call to 'closemedia'.
; 3a. Free all partitions.
	push	esi edi
	mov	edi, [esi+DISK.NumPartitions]
	mov	esi, [esi+DISK.Partitions]
	test	edi, edi
	jz	.nofree
.freeloop:
	lodsd
	call	free
	dec	edi
	jnz	.freeloop
.nofree:
	pop	edi esi
; 3b. Free the cache.
	call	disk_free_cache
; 3c. Call the driver.
	mov	al, DISKFUNC.closemedia
	stdcall disk_call_driver
; 3d. Clear the flag.
	mov	[esi+DISK.MediaUsed], 0
.nothing:
	ret

; This function is called by the driver and informs the kernel that the media
; has changed. If the media is non-removable, it is called exactly once
; immediately after 'disk_add' and once from 'disk_del'.
; Parameters:
; [esp+4] = handle of the disk, i.e. the pointer to the DISK structure.
; [esp+8] = new status of the media: zero = no media, nonzero = media inserted.
disk_media_changed:
	push	ebx esi edi		; save used registers to be stdcall
; 1. Remove the existing media, if it is present.
	mov	esi, [esp+4+12] 	; esi = pointer to DISK
; 1a. Check whether it is present. Since DISK.MediaInserted is changed only
; in this function and calls to this function are synchronized, no lock is
; required for checking.
	cmp	[esi+DISK.MediaInserted], 0
	jz	.noremove
; We really need to remove the media.
; 1b. Acquire mutex.
	lea	ecx, [esi+DISK.MediaLock]
	call	mutex_lock
; 1c. Clear the flag.
	mov	[esi+DISK.MediaInserted], 0
; 1d. Release mutex.
	call	mutex_unlock
; 1e. Remove the "lifetime" reference and possibly destroy the structure.
	call	disk_media_dereference
.noremove:
; 2. Test whether there is new media.
	cmp	dword [esp+8+12], 0
	jz	.noinsert
; Yep, there is.
; 3. Process the new media. We assume that all media fields are available to
; use, see comments in 'disk_media_dereference' (this covers using by previous
; media referencers) and note that calls to this function are synchronized
; (this covers using by new media referencers).
; 3a. Call the 'querymedia' callback.
; .Flags are set to zero for possible future extensions.
	lea	edx, [esi+DISK.MediaInfo]
	and	[edx+DISKMEDIAINFO.Flags], 0
	mov	al, DISKFUNC.querymedia
	stdcall disk_call_driver, edx
; 3b. Check the result of the callback. Abort if it failed.
	test	eax, eax
	jnz	.noinsert
; 3c. Allocate the cache unless disabled by the driver. Abort if failed.
	call	disk_init_cache
	test	al, al
	jz	.noinsert
; 3d. Acquire the lifetime reference for the media object.
	inc	[esi+DISK.MediaRefCount]
; 3e. Scan for partitions. Ignore result; the list of partitions is valid even
; on errors.
	call	disk_scan_partitions
; 3f. Media is inserted and available for use.
	inc	[esi+DISK.MediaInserted]
.noinsert:
; 4. Return.
	pop	edi esi ebx		; restore used registers to be stdcall
	ret	8			; purge 2 dword arguments to be stdcall

; This function is a thunk for all functions of a disk driver.
; It checks whether the referenced function is implemented in the driver.
; If so, this function jumps to the function in the driver.
; Otherwise, it jumps to the default implementation.
; al = offset of function in the DISKFUNC structure;
; esi = pointer to the DISK structure;
; stack is the same as for the corresponding function except that the
; first parameter (void* userdata) is prepended automatically.
disk_call_driver:
	movzx	eax, al ; eax = offset of function in the DISKFUNC structure
; 1. Prepend the first argument to the stack.
	pop	ecx	; ecx = return address
	push	[esi+DISK.UserData]	; add argument
	push	ecx	; save return address
; 2. Check that the required function is inside the table. If not, go to 5.
	mov	ecx, [esi+DISK.Functions]
	cmp	eax, [ecx+DISKFUNC.strucsize]
	jae	.default
; 3. Check that the required function is implemented. If not, go to 5.
	mov	ecx, [ecx+eax]
	test	ecx, ecx
	jz	.default
; 4. Jump to the required function.
	jmp	ecx
.default:
; 5. Driver does not implement the required function; use default implementation.
	jmp	dword [disk_default_callbacks+eax-4]

; The default implementation of DISKFUNC.querymedia.
disk_default_querymedia:
	push	DISK_STATUS_INVALID_CALL
	pop	eax
	ret	8

; The default implementation of DISKFUNC.read and DISKFUNC.write.
disk_default_read:
disk_default_write:
	push	DISK_STATUS_INVALID_CALL
	pop	eax
	ret	20

; The default implementation of DISKFUNC.close, DISKFUNC.closemedia and
; DISKFUNC.flush.
disk_default_close:
disk_default_closemedia:
disk_default_flush:
	xor	eax, eax
	ret	4

; The default implementation of DISKFUNC.adjust_cache_size.
disk_default_adjust_cache_size:
	mov	eax, [esp+4]
	ret	4

; This is an internal function called from 'disk_media_changed' when a new media
; is detected. It creates the list of partitions for the media.
; If media is not partitioned, then the list consists of one partition which
; covers all the media.
; esi = pointer to the DISK structure.
disk_scan_partitions:
; 1. Initialize .NumPartitions and .Partitions fields as zeros: empty list.
	and	[esi+DISK.NumPartitions], 0
	and	[esi+DISK.Partitions], 0
; 2. Currently we can work only with 512-bytes sectors. Check this restriction.
; The only exception is 2048-bytes CD/DVD, but they are not supported yet by
; this code.
	cmp	[esi+DISK.MediaInfo.SectorSize], 512
	jz	.doscan
	DEBUGF 1,'K : sector size is %d, only 512 is supported\n',[esi+DISK.MediaInfo.SectorSize]
	ret
.doscan:
; 3. Acquire the buffer for MBR and bootsector tests. See the comment before
; the 'partition_buffer_users' variable.
	mov	ebx, mbr_buffer 	; assume the global buffer is free
lock	inc	[partition_buffer_users]
	jz	.buffer_acquired	; yes, it is free
lock	dec	[partition_buffer_users]	; no, we must allocate
	stdcall kernel_alloc, 512*3
	test	eax, eax
	jz	.nothing
	xchg	eax, ebx
.buffer_acquired:
; MBR/EBRs are organized in the chain. We use a loop over MBR/EBRs, but no
; more than MAX_NUM_PARTITION times.
; 4. Prepare things for the loop.
; ebp will hold the sector number for current MBR/EBR.
; [esp] will hold the sector number for current extended partition, if there
; is one.
; [esp+4] will hold the counter that prevents long loops.
	push	ebp		; save ebp
	push	MAX_NUM_PARTITIONS	; the counter of max MBRs to process
	xor	ebp, ebp	; start from sector zero
	push	ebp		; no extended partition yet
.new_mbr:
; 5. Read the current sector.
; Note that 'read' callback operates with 64-bit sector numbers, so we must
; push additional zero as a high dword of sector number.
	mov	al, DISKFUNC.read
	push	1
	stdcall disk_call_driver, ebx, ebp, 0, esp
	pop	ecx
; 6. If the read has failed, abort the loop.
	dec	ecx
	jnz	.mbr_failed
; 7. Check the MBR/EBR signature. If it is wrong, abort the loop.
; Soon we will access the partition table which starts at ebx+0x1BE,
; so we can fill its address right now. If we do it now, then the addressing
; [ecx+0x40] is shorter than [ebx+0x1fe]: one-byte offset vs 4-bytes offset.
	lea	ecx, [ebx+0x1be]	; ecx -> partition table
	cmp	word [ecx+0x40], 0xaa55
	jnz	.mbr_failed
; 8. The MBR is treated differently from EBRs. For MBR we additionally need to
; execute step 9 and possibly step 10.
	test	ebp, ebp
	jnz	.mbr
; The partition table can be present or not present. In the first case, we just
; read the MBR. In the second case, we just read the bootsector for a
; filesystem.
; The following algorithm is used to distinguish between these cases.
; A. If at least one entry of the partition table is invalid, this is
;    a bootsector. See the description of 'is_partition_table_entry' for
;    definition of validity.
; B. If all entries are empty (filesystem type field is zero) and the first
;    byte is jmp opcode (0EBh or 0E9h), this is a bootsector which happens to
;    have zeros in the place of partition table.
; C. Otherwise, this is an MBR.
; 9. Test for MBR vs bootsector.
; 9a. Check entries. If any is invalid, go to 10 (rule A).
	call	is_partition_table_entry
	jc	.notmbr
	add	ecx, 10h
	call	is_partition_table_entry
	jc	.notmbr
	add	ecx, 10h
	call	is_partition_table_entry
	jc	.notmbr
	add	ecx, 10h
	call	is_partition_table_entry
	jc	.notmbr
; 9b. Check types of the entries. If at least one is nonzero, go to 11 (rule C).
	mov	al, [ecx-30h+PARTITION_TABLE_ENTRY.Type]
	or	al, [ecx-20h+PARTITION_TABLE_ENTRY.Type]
	or	al, [ecx-10h+PARTITION_TABLE_ENTRY.Type]
	or	al, [ecx+PARTITION_TABLE_ENTRY.Type]
	jnz	.mbr
; 9c. Empty partition table or bootsector with many zeroes? (rule B)
	cmp	byte [ebx], 0EBh
	jz	.notmbr
	cmp	byte [ebx], 0E9h
	jnz	.mbr
.notmbr:
; 10. This is not an  MBR. The media is not partitioned. Create one partition
; which covers all the media and abort the loop.
	stdcall disk_add_partition, 0, 0, \
		dword [esi+DISK.MediaInfo.Capacity], dword [esi+DISK.MediaInfo.Capacity+4]
	jmp	.done
.mbr:
; 11. Process all entries of the new MBR/EBR
	lea	ecx, [ebx+0x1be]	; ecx -> partition table
	push	0	; assume no extended partition
	call	process_partition_table_entry
	add	ecx, 10h
	call	process_partition_table_entry
	add	ecx, 10h
	call	process_partition_table_entry
	add	ecx, 10h
	call	process_partition_table_entry
	pop	ebp
; 12. Test whether we found a new EBR and should continue the loop.
; 12a. If there was no next EBR, return.
	test	ebp, ebp
	jz	.done
; Ok, we have EBR.
; 12b. EBRs addresses are relative to the start of extended partition.
; For simplicity, just abort if an 32-bit overflow occurs; large disks
; are most likely partitioned with GPT, not MBR scheme, since the precise
; calculation here would increase limit just twice at the price of big
; compatibility problems.
	pop	eax	; load extended partition
	add	ebp, eax
	jc	.mbr_failed
; 12c. If extended partition has not yet started, start it.
	test	eax, eax
	jnz	@f
	mov	eax, ebp
@@:
; 12c. If the limit is not exceeded, continue the loop.
	dec	dword [esp]
	push	eax	; store extended partition
	jnz	.new_mbr
.mbr_failed:
.done:
; 13. Cleanup after the loop.
	pop	eax	; not important anymore
	pop	eax	; not important anymore
	pop	ebp	; restore ebp
; 14. Release the buffer.
; 14a. Test whether it is the global buffer or we have allocated it.
	cmp	ebx, mbr_buffer
	jz	.release_partition_buffer
; 14b. If we have allocated it, free it.
	xchg	eax, ebx
	call	free
	jmp	.nothing
; 14c. Otherwise, release reference.
.release_partition_buffer:
lock	dec	[partition_buffer_users]
.nothing:
; 15. Return.
	ret

; This is an internal function called from disk_scan_partitions. It checks
; whether the entry pointed to by ecx is a valid entry of partition table.
; The entry is valid if the first byte is 0 or 80h, the first sector plus the
; length is less than twice the size of media. Multiplication by two is
; required since the size mentioned in the partition table can be slightly
; greater than the real size.
is_partition_table_entry:
; 1. Check .Bootable field.
	mov	al, [ecx+PARTITION_TABLE_ENTRY.Bootable]
	and	al, 7Fh
	jnz	.invalid
; 3. Calculate first sector + length. Note that .FirstAbsSector is relative
; to the MBR/EBR, so the real sum is ebp + .FirstAbsSector + .Length.
	mov	eax, ebp
	xor	edx, edx
	add	eax, [ecx+PARTITION_TABLE_ENTRY.FirstAbsSector]
	adc	edx, 0
	add	eax, [ecx+PARTITION_TABLE_ENTRY.Length]
	adc	edx, 0
; 4. Divide by two.
	shr	edx, 1
	rcr	eax, 1
; 5. Compare with capacity. If the subtraction (edx:eax) - .Capacity does not
; overflow, this is bad.
	sub	eax, dword [esi+DISK.MediaInfo.Capacity]
	sbb	edx, dword [esi+DISK.MediaInfo.Capacity+4]
	jnc	.invalid
.valid:
; 5. Return success: CF is cleared.
	clc
	ret
.invalid:
; 6. Return fail: CF is set.
	stc
	ret

; This is an internal function called from disk_scan_partitions. It processes
; the entry pointed to by ecx.
; * If the entry is invalid, just ignore this entry.
; * If the type is zero, just ignore this entry.
; * If the type is one of types for extended partition, store the address
;   of this partition as the new MBR in [esp+4].
; * Otherwise, add the partition to the list of partitions for this disk.
;   We don't use the type from the entry to identify the file system;
;   fs-specific checks do this more reliably.
process_partition_table_entry:
; 1. Check for valid entry. If invalid, return (go to 5).
	call	is_partition_table_entry
	jc	.nothing
; 2. Check for empty entry. If invalid, return (go to 5).
	mov	al, [ecx+PARTITION_TABLE_ENTRY.Type]
	test	al, al
	jz	.nothing
; 3. Check for extended partition. If extended, go to 6.
irp type,\
    0x05,\		   ; DOS: extended partition
    0x0f,\		   ; WIN95: extended partition, LBA-mapped
    0xc5,\		   ; DRDOS/secured: extended partition
    0xd5		   ; Old Multiuser DOS secured: extended partition
{
	cmp	al, type
	jz	.extended
}
; 4. If we are here, that is a normal partition. Add it to the list.
; Note that the first sector is relative to MBR/EBR.
	mov	eax, ebp
	xor	edx, edx
	add	eax, [ecx+PARTITION_TABLE_ENTRY.FirstAbsSector]
	adc	edx, 0
	push	ecx
	stdcall disk_add_partition, eax, edx, \
		[ecx+PARTITION_TABLE_ENTRY.Length], 0
	pop	ecx
.nothing:
; 5. Return.
	ret
.extended:
; 6. If we are here, that is an extended partition. Store the address.
	mov	eax, [ecx+PARTITION_TABLE_ENTRY.FirstAbsSector]
	mov	[esp+4], eax
	ret

; This is an internal function called from disk_scan_partitions and
; process_partition_table_entry. It adds one partition to the list of
; partitions for the media.
proc disk_add_partition stdcall uses ebx edi, start:qword, length:qword
; 1. Check that this partition will not exceed the limit on total number.
	cmp	[esi+DISK.NumPartitions], MAX_NUM_PARTITIONS
	jae	.nothing
; 2. Check that this partition does not overlap with any already registered
; partition. Since any file system assumes that the disk data will not change
; outside of its control, such overlap could be destructive.
; Since the number of partitions is usually very small and is guaranteed not
; to be large, the simple linear search is sufficient.
; 2a. Prepare the loop: edi will point to the current item of .Partitions
; array, ecx will be the current item, ebx will hold number of items left.
	mov	edi, [esi+DISK.Partitions]
	mov	ebx, [esi+DISK.NumPartitions]
	test	ebx, ebx
	jz	.partitionok
.scan_existing:
; 2b. Get the next partition.
	mov	ecx, [edi]
	add	edi, 4
; The range [.FirstSector, .FirstSector+.Length) must be either entirely to
; the left of [start, start+length) or entirely to the right.
; 2c. Subtract .FirstSector - start. The possible overflow distinguish between
; cases "to the left" (2e) and "to the right" (2d).
	mov	eax, dword [ecx+PARTITION.FirstSector]
	mov	edx, dword [ecx+PARTITION.FirstSector+4]
	sub	eax, dword [start]
	sbb	edx, dword [start+4]
	jb	.less
; 2d. .FirstSector is greater than or equal to start. Check that .FirstSector
; is greater than or equal to start+length; the subtraction
; (.FirstSector-start) - length must not cause overflow. Go to 2g if life is
; good or to 2f in the other case.
	sub	eax, dword [length]
	sbb	edx, dword [length+4]
	jb	.overlap
	jmp	.next_existing
.less:
; 2e. .FirstSector is less than start. Check that .FirstSector+.Length is less
; than or equal to start. If the addition (.FirstSector-start) + .Length does
; not cause overflow, then .FirstSector + .Length is strictly less than start;
; since the equality is also valid, use decrement preliminarily. Go to 2g or
; 2f depending on the overflow.
	sub	eax, 1
	sbb	edx, 0
	add	eax, dword [ecx+PARTITION.Length]
	adc	edx, dword [ecx+PARTITION.Length+4]
	jnc	.next_existing
.overlap:
; 2f. The partition overlaps with previously registered partition. Say warning
; and return with nothing done.
	dbgstr 'two partitions overlap, ignoring the last one'
	jmp	.nothing
.next_existing:
; 2g. The partition does not overlap with the current partition. Continue the
; loop.
	dec	ebx
	jnz	.scan_existing
.partitionok:
; 3. The partition has passed tests. Reallocate the partitions array for a new
; entry.
; 3a. Call the allocator.
	mov	eax, [esi+DISK.NumPartitions]
	inc	eax	; one more entry
	shl	eax, 2	; each entry is dword
	call	malloc
; 3b. Test the result. If failed, return with nothing done.
	test	eax, eax
	jz	.nothing
; 3c. Copy the old array to the new array.
	mov	edi, eax
	push	esi
	mov	ecx, [esi+DISK.NumPartitions]
	mov	esi, [esi+DISK.Partitions]
	rep	movsd
	pop	esi
; 3d. Set the field in the DISK structure to the new array.
	xchg	[esi+DISK.Partitions], eax
; 3e. Free the old array.
	call	free
; 4. Recognize the file system.
; 4a. Call the filesystem recognizer. It will allocate the PARTITION structure
; with possible filesystem-specific fields.
	call	disk_detect_partition
; 4b. Check return value. If zero, return with list not changed; so far only
; the array was reallocated, this is ok for other code.
	test	eax, eax
	jz	.nothing
; 5. Insert the new partition to the list.
	stosd
	inc	[esi+DISK.NumPartitions]
; 6. Return.
.nothing:
	ret
endp

; This is an internal function called from disk_add_partition.
; It tries to recognize the file system on the partition and allocates the
; corresponding PARTITION structure with filesystem-specific fields.
disk_detect_partition:
; This function inherits the stack frame from disk_add_partition. In stdcall
; with ebp-based frame arguments start from ebp+8, since [ebp]=saved ebp
; and [ebp+4]=return address.
virtual at ebp+8
.start	dq	?
.length dq	?
end virtual
; Currently no file systems are supported, so just allocate the PARTITION
; structure without extra fields.
; 1. Allocate and check result.
	push	sizeof.PARTITION
	pop	eax
	call	malloc
	test	eax, eax
	jz	.nothing
; 2. Fill the common fields: copy .start and .length.
	mov	edx, dword [.start]
	mov	dword [eax+PARTITION.FirstSector], edx
	mov	edx, dword [.start+4]
	mov	dword [eax+PARTITION.FirstSector+4], edx
	mov	edx, dword [.length]
	mov	dword [eax+PARTITION.Length], edx
	mov	edx, dword [.length+4]
	mov	dword [eax+PARTITION.Length+4], edx
.nothing:
; 3. Return with eax = pointer to PARTITION or NULL.
	ret

; This function is called from file_system_lfn.
; This handler gets the control each time when fn 70 is called
; with unknown item of root subdirectory.
; in: esi -> name
;     ebp = 0 or rest of name relative to esi
; out: if the handler processes path, it must not return in file_system_lfn,
;      but instead pop return address and return directly to the caller
;      otherwise simply return
dyndisk_handler:
	push	ebx edi 	; save registers used in file_system_lfn
; 1. Acquire the mutex.
	mov	ecx, disk_list_mutex
	call	mutex_lock
; 2. Loop over the list of DISK structures.
; 2a. Initialize.
	mov	ebx, disk_list
.scan:
; 2b. Get the next item.
	mov	ebx, [ebx+DISK.Next]
; 2c. Check whether the list is done. If so, go to 3.
	cmp	ebx, disk_list
	jz	.notfound
; 2d. Compare names. If names match, go to 5.
	mov	edi, [ebx+DISK.Name]
	push	esi
@@:
; esi points to the name from fs operation; it is terminated by zero or slash.
	lodsb
	test	al, al
	jz	.eoin_dec
	cmp	al, '/'
	jz	.eoin
; edi points to the disk name.
	inc	edi
; edi points to lowercase name, this is a requirement for the driver.
; Characters at esi can have any register. Lowercase the current character.
; This lowercasing works for latin letters and digits; since the disk name
; should not contain other symbols, this is ok.
	or	al, 20h
	cmp	al, [edi-1]
	jz	@b
.wrongname:
; 2f. Names don't match. Continue the loop.
	pop	esi
	jmp	.scan
.notfound:
; The loop is done and no name matches.
; 3. Release the mutex.
	call mutex_unlock
; 4. Return normally.
	pop	edi ebx 	; restore registers used in file_system_lfn
	ret
; part of 2d: the name matches partially, but we must check that this is full
; equality.
.eoin_dec:
	dec	esi
.eoin:
	cmp	byte [edi], 0
	jnz	.wrongname
; We found the addressed DISK structure.
; 5. Reference the disk.
lock	inc	[ebx+DISK.RefCount]
; 6. Now we are sure that the DISK structure is not going to die at least
; while we are working with it, so release the global mutex.
	call	mutex_unlock
; 7. Acquire the mutex for media object.
	pop	edi		; restore edi
	lea	ecx, [ebx+DISK.MediaLock]
	call	mutex_lock
; 8. Get the media object. If it is not NULL, reference it.
	xor	edx, edx
	cmp	[ebx+DISK.MediaInserted], dl
	jz	@f
	mov	edx, ebx
	inc	[ebx+DISK.MediaRefCount]
@@:
; 9. Now we are sure that the media object, if it exists, is not going to die
; at least while we are working with it, so release the mutex for media object.
	call	mutex_unlock
	mov	ecx, ebx
	pop	ebx eax 	; restore ebx, pop return address
; 10. Check whether the fs operation wants to enumerate partitions (go to 11)
; or work with some concrete partition (go to 12).
	cmp	byte [esi], 0
	jnz	.haspartition
; 11. The fs operation wants to enumerate partitions.
; 11a. Only "list directory" operation is applicable to /<diskname> path. Check
; the operation code. If wrong, go to 13.
	cmp	dword [ebx], 1
	jnz	.access_denied
; 11b. If the media is inserted, use 'fs_dyndisk_next' as an enumeration
; procedure. Otherwise, use 'fs_dyndisk_next_nomedia'.
	mov	esi, fs_dyndisk_next_nomedia
	test	edx, edx
	jz	@f
	mov	esi, fs_dyndisk_next
@@:
; 11c. Let the procedure from fs_lfn.inc do the job.
	jmp	file_system_lfn.maindir_noesi
.haspartition:
; 12. The fs operation has specified some partition.
; 12a. Store parameters for callback functions.
	push	edx
	push	ecx
; 12b. Store callback functions.
	push	dyndisk_cleanup
	push	fs_dyndisk
	mov	edi, esp
; 12c. Let the procedure from fs_lfn.inc do the job.
	jmp	file_system_lfn.found2
.access_denied:
; 13. Fail the operation with the appropriate code.
	mov	dword [esp+32], ERROR_ACCESS_DENIED
.cleanup:
; 14. Cleanup.
	mov	esi, ecx	; disk*dereference assume that esi points to DISK
.cleanup_esi:
	test	edx, edx	; if there are no media, we didn't reference it
	jz	@f
	call	disk_media_dereference
@@:
	call	disk_dereference
; 15. Return.
	ret

; This is a callback for cleaning up things called from file_system_lfn.found2.
dyndisk_cleanup:
	mov	esi, [edi+8]
	mov	edx, [edi+12]
	jmp	dyndisk_handler.cleanup_esi

; This is a callback for enumerating partitions called from
; file_system_lfn.maindir in the case of inserted media.
; It just increments eax until DISK.NumPartitions reached and then
; cleans up.
fs_dyndisk_next:
	cmp	eax, [ecx+DISK.NumPartitions]
	jae	.nomore
	inc	eax
	clc
	ret
.nomore:
	pusha
	mov	esi, ecx
	call	disk_media_dereference
	call	disk_dereference
	popa
	stc
	ret

; This is a callback for enumerating partitions called from
; file_system_lfn.maindir in the case of missing media.
; In this case we create one pseudo-partition.
fs_dyndisk_next_nomedia:
	cmp	eax, 1
	jae	.nomore
	inc	eax
	clc
	ret
.nomore:
	pusha
	mov	esi, ecx
	call	disk_dereference
	popa
	stc
	ret

; This is a callback for doing real work with selected partition.
; Currently this is just placeholder, since no file systems are supported.
; edi = esp -> {dd fs_dyndisk, dd dyndisk_cleanup, dd pointer to DISK, dd media object}
; ecx = partition number, esi+ebp = ASCIIZ name
fs_dyndisk:
	dec	ecx	; convert to zero-based partition index
	pop	edx edx edx eax ; edx = pointer to DISK, eax = NULL or edx
	test	eax, eax
	jz	.nomedia
.main:
	cmp	ecx, [edx+DISK.NumPartitions]
	jae	.notfound
	mov	dword [esp+32], ERROR_UNKNOWN_FS
.cleanup:
	mov	esi, edx
	call	disk_media_dereference
	call	disk_dereference
	ret
.notfound:
	mov	dword [esp+32], ERROR_FILE_NOT_FOUND
	jmp	.cleanup
.nomedia:
	test	ecx, ecx
	jnz	.notfound
	test	byte [edx+DISK.DriverFlags], DISK_NO_INSERT_NOTIFICATION
	jz	.deverror
; if the driver does not support insert notifications and we are the only fs
; operation with this disk, issue the fake insert notification; if media is
; still not inserted, 'disk_media_changed' will detect this and do nothing
;;;        push    ebx
	lea	ecx, [edx+DISK.MediaLock]
	call	mutex_lock
	cmp	[edx+DISK.MediaRefCount], 1
	jnz	.noluck
	call	mutex_unlock
	push	edx
	stdcall disk_media_changed, edx, 1
	pop	edx
	lea	ecx, [edx+DISK.MediaLock]
	call	mutex_lock
	cmp	[edx+DISK.MediaInserted], 0
	jz	.noluck
lock	inc	[edx+DISK.MediaRefCount]
	call	mutex_unlock
	xor	ecx, ecx
	jmp	.main
.noluck:
	call	mutex_unlock
.deverror:
	mov	dword [esp+32], ERROR_DEVICE
	mov	esi, edx
	call	disk_dereference
	ret

; This function is called from file_system_lfn.
; This handler is called when virtual root is enumerated
; and must return all items which can be handled by this.
; It is called several times, first time with eax=0
; in: eax = 0 for first call, previously returned value for subsequent calls
; out: eax = 0 => no more items
;      eax != 0 => buffer pointed to by edi contains name of item
dyndisk_enum_root:
	push	edx		; save register used in file_system_lfn
	mov	ecx, disk_list_mutex	; it will be useful
; 1. If this is the first call, acquire the mutex and initialize.
	test	eax, eax
	jnz	.notfirst
	call	mutex_lock
	mov	eax, disk_list
.notfirst:
; 2. Get next item.
	mov	eax, [eax+DISK.Next]
; 3. If there are no more items, go to 6.
	cmp	eax, disk_list
	jz	.last
; 4. Copy name from the DISK structure to edi.
	push	eax esi
	mov	esi, [eax+DISK.Name]
@@:
	lodsb
	stosb
	test	al, al
	jnz	@b
	pop	esi eax
; 5. Return with eax = item.
	pop	edx		; restore register used in file_system_lfn
	ret
.last:
; 6. Release the mutex and return with eax = 0.
	call	mutex_unlock
	xor	eax, eax
	pop	edx		; restore register used in file_system_lfn
	ret