Module Plasma_filebuf

module Plasma_filebuf: sig .. end

type errno = Plasma_util.errno 

type strmem = [ `Memory of Netsys_mem.memory | `String of string ] 

type buf_state = [ `Clean
       | `Dirty
       | `Dropped
       | `Invalid
       | `Reading of int64 option Uq_engines.engine
       | `Writing of int64 option Uq_engines.engine
       | `Written ] 

Buffer states:

• an `Invalid buffer is reserved for a certain block but it is not yet filled with any meaningful data
• a `Clean buffer contains valid data of a file which is unmodified
• a `Dirty buffer contains modified data of a file (which still needs to be written)
• a `Reading buffer is being filled with data. For the user this is very much like `Invalid (data are unusable), but the argument engine terminates when the buffer changes state again.
• a `Writing buffer is being written out. For the user this is very much like `Clean (data can be read but not modified again). The argument engine terminates when the buffer changes state again.
• the `Written state is used after `Writing as long as it is still unclear whether the write is successful or not. On success, the buffer can be set to `Clean again. On error, it will go back to `Dirty.
• a `Dropped buffer is reused for a different purpose

Buffer descriptor. The user of this API should never modify entries of it.

A buffer descriptor is handed out for the lifetime of a buffer. The states are:

• A freshly designated buffer is in `Invalid state
• While reading the block the buffer is in `Reading state
• If `Reading is successful and there is no pressure to reassign the buffer immediately, it becomes `Clean. A `Clean buffer can be dropped by the system at any time.
• When the contents are modified the buffer becomes `Dirty
• Dirty buffers can be written to disk. The buffer enters `Writing state. Note that it is not allowed to modify the buffer while in `Writing state - one must wait for the completion of the write first
• A written buffer is set to `Written. This is a special state meaning that the write is done but not yet committed.
• The contents of a written buffer can be modified. This is only recorded by the buf_dirty flag.
• After the commit of the write, the `Written buffer becomes `Clean or `Dirty, depending on the argument flag. If the commit is not successful the buffer is set to `Dirty.

When a clean buffer is dropped, the state in the descriptor is set to `Dropped. At the same time, the buffer memory is reused for a different descriptor.

The `Invalid state must not be kept for longer than a moment. If there are several requests for the same block, and the buffer is `Invalid the other requests can only use busy waiting to handle this case.

The buf_dirty flag can be set in some contexts to indicate that there was a data modification and `Dirty should be entered next.

type buffer_system 

val create_buffer_system : int -> int -> Unixqueue.event_system -> buffer_system

create_buffer_system n_bufs blocksize esys

val mem_size : buffer_system -> int

size of buffers in bytes

val blit_from_buffer : buffer_system ->
       buffer -> int -> strmem -> int -> int -> unit

val blit_to_buffer : buffer_system ->
       strmem -> int -> buffer -> int -> int -> unit

val clean_access : buffer_system -> buffer -> unit

a clean buffer is continued to be used as clean buffer after a read access

val dirty_access : buffer_system -> buffer -> int64 -> unit

A clean or invalid buffer is set to dirty. The int is the EOF position that should now be ensured at least

val switch_to_reading : buffer_system ->
       buffer ->
       int64 option Uq_engines.engine -> int64 option -> unit

switch_to_reading sys b e eof_opt:

The buffer is switched to `Reading state (from either `Invalid, `Clean or `Dirty). The engine e must be in a non-final state. When the engine terminates, this is taken as indication that the read is finished. The return value is seqno_opt. If Some seqno the read has been successful and the buffer is filled with data of this sequence number of the file. The return value None means error. This function arranges that the buffer is switched to a follow-up state when the engine is finished:

• if the buffer is scheduled for being dropped, this is done now
• the buffer is also dropped when the read is non-successful
• if eof_opt <> None the buffer becomes `Dirty, and the EOF value is considered for the flush request
• if there is pressure for memory the buffer is reassigned
• otherwise the buffer becomes `Clean

val switch_to_writing : buffer_system ->
       buffer -> int64 option Uq_engines.engine -> unit

The buffer is switched to `Writing state (from `Dirty). The argument engine must be in a non-final state. When the engine terminates, this is taken as indication that the buffer is written out. This function arranges that the buffer state is switched again when this occurs. The follow-up state is always `Written.

The engine returns the new sequence number.

A `Written buffer should be committed, and then one of three functions needs to be called:

• write_committed: if the commit has been successful
• write_erroneous: if an error occurred
• write_cancelled: if the block will no longer be written because there was an error for another block of the same transaction

val schedule_drop : buffer_system -> buffer -> unit

If the buffer can be immediately invalidated this is done. Otherwise the buf_delayed_drop flag is set, and and the end of the ongoing read/write the buffer will be set to `Dropped.

val schedule_drop_inode : buffer_system -> int64 -> unit

Same for a whole inode. It is no error if there is no buffer for this inode

val lookup_buffer : buffer_system -> int64 -> int64 -> buffer

lookup_buffer sys inode index: Looks the buffer up for inode and index, or raises Not_found if there is none yet, or the existing descriptor is in `Dropped state.

val request_buffer_e : buffer_system ->
       int64 -> int64 -> buffer option Uq_engines.engine

request_buffer_e sys inode index: This function is to be used when lookup_buffer raises Not_found to get a new buffer. The new buffer is initially in `Invalid state. Note that competing engine-driven threads can see this state in a certain moment.

Note that one should immediately set the buffer to a different state when the engine is done.

The function returns None via the engine when the buffer already exists or when a concurrent request was faster (if there are several calls of request_buffer_e for the same inode/index pair, only one call gets the buffer, and the other calls see None). The function also returns None when an error is recorded for this inode (with write_erroneous). Because of the latter, it is recommended to reset_inode_error just before requesting a buffer.

val select_for_flush : buffer_system -> int64 -> flush_request

select_for_flush sys inode: Returns a list of blocks that need to be written. Once a block is returned by select_for_flush it is not again returned in future calls of this function unless it is set to `Dirty again after leaving the `Written state.

The buf_flushing flag is used for managing this. If buf_flushing it is prevented that a call of dirty_access records the block again for flushing. This flag is cleared when entering `Writing.

This function may raise Not_found if nothing appropriate is found.

val select_inodes : buffer_system -> int64 list

Returns the inodes that can be flushed, in the order of precedence.

val get_flush_min_eof : buffer_system -> int64 -> int64

val get_flush_min_mtime : buffer_system -> int64 -> Plasma_rpcapi_aux.time

Return the values from the flush record (or Not_found)

val write_committed : buffer_system -> buffer -> unit

Records that the write is committed: If the buffer is in `Written false state it is reset to `Clean (or reassigned for a different purpose). If it is in `Written true state it is set to `Dirty.

It is an error to call this function for a different state.

val write_cancelled : buffer_system -> buffer -> unit

The write is cancelled. The buffer becomes dirty again

val write_erroneous : buffer_system -> int64 -> errno -> unit

Records a write error. This should be called after write_cancelled with the error code.

Any pending request_buffer_e for this inode is interrupted, so None is returned.

val release_flush_request : buffer_system -> flush_request -> unit

Checks whether the buffers are in the right state after finishing or aborting a flush request. In particular, all buffers are set to `Dirty that are still in `Written state. Also, the buf_flushing flag is cleared.

val inode_error : buffer_system -> int64 -> errno option

Whether there is a write error for this inode

val reset_inode_error : buffer_system -> int64 -> unit

Resets the error

val n_waiting : buffer_system -> int

The number of buffer requests that can be currently not satisfied

val dump_buffers : buffer_system -> unit

Writes the buffer table to log (debug level)

val max_time : Plasma_rpcapi_aux.time -> Plasma_rpcapi_aux.time -> Plasma_rpcapi_aux.time

The max of two time structs