I would appreciate any donations. Wishlist or send e-mail type donations to maekawa AT daemon-systems.org.
PUFFS_FRAMEBUF(3) Library Functions Manual PUFFS_FRAMEBUF(3) NAME puffs_framebuf - buffering and event handling for networked file systems LIBRARY puffs Convenience Library (libpuffs, -lpuffs) SYNOPSIS #include <puffs.h> struct puffs_framebuf * puffs_framebuf_make(); void puffs_framebuf_destroy(struct puffs_framebuf *pufbuf); void puffs_framebuf_recycle(struct puffs_framebuf *pufbuf); int puffs_framebuf_reserve_space(struct puffs_framebuf *pufbuf, size_t space); int puffs_framebuf_putdata(struct puffs_framebuf *pufbuf, const void *data, size_t dlen); int puffs_framebuf_putdata_atoff(struct puffs_framebuf *pufbuf, size_t offset, const void *data, size_t dlen); int puffs_framebuf_getdata(struct puffs_framebuf *pufbuf, void *data, size_t dlen); int puffs_framebuf_getdata_atoff(struct puffs_framebuf *pufbuf, size_t offset, void *data, size_t dlen); size_t puffs_framebuf_telloff(struct puffs_framebuf *pufbuf); size_t puffs_framebuf_tellsize(struct puffs_framebuf *pufbuf); size_t puffs_framebuf_remaining(struct puffs_framebuf *pufbuf); int puffs_framebuf_seekset(struct puffs_framebuf *pufbuf, size_t offset); int puffs_framebuf_getwindow(struct puffs_framebuf *pufbuf, size_t offset, void **winp, size_t *winlen); int puffs_framev_enqueue_cc(struct puffs_cc *pcc, int fd, struct puffs_framebuf *pufbuf, int flags); void puffs_framev_cb(struct puffs_usermount *pu, int fd, struct puffs_framebuf *pufbuf, void *arg, int flags, int error); void puffs_framev_enqueue_cb(struct puffs_usermount *pu, int fd, struct puffs_framebuf *pufbuf, puffs_framebuf_cb fcb, void *fcb_arg, int flags); void puffs_framev_enqueue_justsend(struct puffs_usermount *pu, int fd, struct puffs_framebuf *pufbuf, int waitreply, int flags); void puffs_framev_enqueue_directsend(struct puffs_usermount *pu, int fd, struct puffs_framebuf *pufbuf, int flags); void puffs_framev_enqueue_directreceive(struct puffs_usermount *pu, int fd, struct puffs_framebuf *pufbuf, int flags); int puffs_framev_framebuf_ccpromote(struct puffs_framebuf *pufbuf, struct puffs_cc *pcc); int puffs_framev_enqueue_waitevent(struct puffs_cc *pcc, int fd, int *what); int puffs_framev_readframe_fn(struct puffs_usermount *pu, struct puffs_framebuf *pufbuf, int fd, int *done); int puffs_framev_writeframe_fn(struct puffs_usermount *pu, struct puffs_framebuf *pufbuf, int fd, int *done); int puffs_framev_cmpframe_fn(struct puffs_usermount *pu, struct puffs_framebuf *cmp1, struct puffs_framebuf *cmp2, int *notresp); void puffs_framev_gotframe_fn(struct puffs_usermount *pu, struct puffs_framebuf *pufbuf); void puffs_framev_fdnotify_fn(struct puffs_usermount *pu, int fd, int what); void puffs_framev_init(struct puffs_usermount *pu, puffs_framev_readframe_fn rfb, puffs_framev_writeframe_fn wfb, puffs_framev_cmpframe_fn cmpfb, puffs_framev_gotframe_fn gotfb, puffs_framev_fdnotify_fn fdnotfn); int puffs_framev_addfd(struct puffs_usermount *pu, int fd, int what); int puffs_framev_enablefd(struct puffs_usermount *pu, int fd, int what); int puffs_framev_disablefd(struct puffs_usermount *pu, int fd, int what); int puffs_framev_removefd(struct puffs_usermount *pu, int fd, int error); void puffs_framev_unmountonclose(struct puffs_usermount *pu, int fd, int what); DESCRIPTION The puffs_framebuf routines provide buffering and an event loop structured around the buffers. It operates on top of the puffs continuation framework, puffs_cc(3), and multiplexes execution automatically to an instance whenever one is runnable. The file system is entered in three different ways: ⊕ An event arrives from the kernel and the puffs_ops(3) callbacks are called to start processing the event. ⊕ A file system which has sent out a request receives a response. Execution is resumed from the place where the file system yielded. ⊕ A request from a peer arrives. A request is an incoming PDU which is not a response to any outstanding request. puffs_framebuf is used by defining various callbacks and providing I/O descriptors, which are then monitored for activity by the library. A descriptor, when present, can be either enabled or disabled for input and output. If a descriptor is not enabled for a certain direction, the callbacks will not be called even if there were activity on the descriptor. For example, even if a network socket has been added and there is input data in the socket buffer, the read callback will be called only if the socket has been enabled for reading. File descriptors are treated like sockets: they have two sides, a read side and a write side. The framework determines that one side of the descriptor has been closed if the supplied I/O callbacks return an error or if the I/O multiplexing call says a side has been closed. It is still possible, from the framework perspective, to write to a file descriptor whose read side is closed. However, it is not possible to wait for a response on such a file descriptor. Conversely, it is possible to read responses from a descriptor whose write side is closed. It should be stressed that the implementation underlying the file descriptor might not support this. The following callbacks can be defined, cf. puffs_framev_init(), and all are optional. None of them should block, because this would cause the entire file server to block. One option is to make the descriptors non- blocking before adding them. rfb Read a frame from the file descriptor onto the specified buffer. wfb Write a frame from the specified buffer into the file descriptor. cmpfb Identify if a buffer is the response to the specified buffer. gotfb Called iff no outstanding request matches the incoming frame. In other words, this is called when we receive a request from a peer. fdnotfn Receive notifications about a change-of-state in a file descriptor's status. Better descriptions for each callback are given below. The buffers of puffs_framebuf provide automatic memory management of buffers for the file servers. They provide a cursor to the current buffer offset. Reading or writing data through the normal routines will advance that cursor. Additionally, the buffer size is provided to the user. It represents the maximum offset where data was written. Generally the write functions will fail if the cannot allocate enough memory to satisfy the buffer length requirements. Read functions will fail if the amount of data written to the buffer is not large enough to satisfy the read. puffs_framebuf_make() Create a buffer. Return the address of the buffer or NULL in case no memory was available. puffs_framebuf_destroy(pufbuf) Free memory used by buffer. puffs_framebuf_recycle(pufbuf) Reset offsets so that buffer can be reused. Does not free memory or reallocate memory. puffs_framebuf_reserve_space(pufbuf, space) Make sure that the buffer has space bytes of available memory starting from the current offset. This is not strictly necessary, but can be used for optimizations where it is known in advance how much memory will be required. puffs_framebuf_putdata(pufbuf, data, dlen) Write dlen amount of data from the address data into the buffer. Moves the offset cursor forward dlen bytes. puffs_framebuf_putdata_atoff(pufbuf, offset, data, dlen) Like puffs_framebuf_putdata(), except writes data at buffer offset offset. It is legal to write past the current end of the buffer. Does NOT modify the current offset cursor. puffs_framebuf_getdata(pufbuf, data, dlen) Read dlen bytes of data from the buffer into data. Advances the offset cursor. puffs_framebuf_getdata_atoff(pufbuf, offset, data, dlen) Read data from buffer position offset. Does NOT modify the offset cursor. puffs_framebuf_telloff(pufbuf) Return the offset into the buffer. puffs_framebuf_tellsize(pufbuf) Return the maximum offset where data has been written, i.e. buffer size. puffs_framebuf_remaining(pufbuf) Distance from current offset to the end of the buffer, i.e. size- offset. puffs_framebuf_seekset(pufbuf, offset) Set the offset cursor to the position offset. This does NOT modify the buffer size, but reserves at least enough memory memory for a write to offset and will fail if memory cannot be allocated. puffs_framebuf_getwindow(pufbuf, offset, winp, winlen) Get a direct memory window into the buffer starting from offset. The maximum mapped window size will be winlen bytes, but this routine might return a smaller window and the caller should always check the actual mapped size after the call. The window is returned in winp. This function not modify the buffer offset, but it DOES set the buffer size to offset + winlen in case that value is greater than the current size. The window is valid until the next until the next puffs_framebuf() call operating on the buffer in question. puffs_framev_enqueue_cc(pcc, fd, pufbuf, flags) Add the buffer pufbuf to outgoing queue of descriptor fd and yield with the continuation pcc. Execution is resumed once a response is received. Returns 0 if the buffer was successfully enqueued (not necessarily delivered) and non-zero to signal a non-recoverable error. Usually the buffer is placed at the end of the output queue. However, if flags contains PUFFS_FBQUEUE_URGENT, pufbuf is placed in the front of the queue to be sent immediately after the current PDU (if any) has been sent. puffs_framev_enqueue_cb(pu, fd, pufbuf, fcb, fcb_arg, flags) Enqueue the buffer pufbuf for outgoing data and immediately return. Once a response arrives, the callback fcb() will be called with the argument fcb_arg. The callback function fcb() is responsible for freeing the buffer. Returns 0 if the buffer was successfully enqueued (not necessarily delivered) and non-zero to signal a non- recoverable error. See puffs_framev_enqueue_cc() for flags. puffs_framev_cb(pu, pufbuf, arg, error) Callback function. Called when a response to a specific request arrives from the server. If error is non-zero, the framework was unable to obtain a response and the function should not examine the contents of pufbuf, only do resource cleanup. May not block. puffs_framev_enqueue_justsend(pu, fd, pufbuf, waitreply, flags) Enqueue the buffer pufbuf for outgoing traffic and immediately return. The parameter waitreply can be used to control if the buffer is to be freed immediately after sending of if a response is expected and the buffer should be freed only after the response arrives (receiving an unexpected message from the server is treated as an error). Returns 0 if the buffer was successfully enqueued (not necessarily delivered) and non-zero to signal a non- recoverable error. See puffs_framev_enqueue_cc() for flags. puffs_framev_enqueue_directsend(pcc, fd, pufbuf, flags) Acts like puffs_framev_enqueue_justsend() with the exception that the call yields until the frame has been sent. As opposed to puffs_framev_enqueue_cc(), the routine does not wait for input, but returns immediately after sending the frame. See puffs_framev_enqueue_cc() for flags. puffs_framev_enqueue_directreceive(pcc, fd, pufbuf, flags) Receive data into pufbuf. This routine yields until a complete frame has been read into the buffer by the readframe routine. See puffs_framev_enqueue_cc() for flags. puffs_framev_framebuf_ccpromote(pufbuf, pcc) Promote the framebuffer pufbuf sent with puffs_framev_enqueue_cb() or puffs_framev_enqueue_justsend() to a wait using pcc and yield until the result arrives. The response from the file server for pufbuf must not yet have arrived. If sent with puffs_framev_enqueue_justsend(), the call must be expecting a response. puffs_framev_enqueue_waitevent(pcc, fd, what) Waits for an event in what to happen on file descriptor fd. The events which happened are returned back in what. The possible events are PUFFS_FBIO_READ, PUFFS_FBIO_WRITE, and PUFFS_FBIO_ERROR, specifying read, write and error conditions, respectively. Error is always checked. This call does not depend on if the events were previously enabled on the file descriptor - the specified events are always checked regardless. There is currently no other way to cancel or timeout a call except by removing the file descriptor in question. This may change in the future. puffs_framev_readframe_fn(pu, pufbuf, fd, done) Callback function. Read at most one frame from file descriptor fd into the buffer pufbuf. If a complete frame is read, the value pointed to by done must be set to 1. This function should return 0 on success (even if a complete frame was not yet read) and a non- zero errno to signal a fatal error. In case a fatal error is returned, the read side of the file descriptor is marked closed. This routine will be called with the same buffer argument until a complete frame has been read. May not block. puffs_framev_writeframe_fn(pu, pufbuf, fd, done) Write the frame contained in pufbuf to the file descriptor fd. In case the entire frame is successfully written, *done should be set to 1. This function should return 0 on success (even if a complete frame was not yet written) and a non-zero errno to signal a fatal error. In case a fatal error is returned, the write side of the file descriptor is marked closed. This routine will be called with the same buffer argument until the complete frame has been written. May not block. It is a good idea to make sure that this function can handle a possible SIGPIPE caused by a closed connection. For example, the file server can opt to trap SIGPIPE or, if writing to a socket, call send() with the flag MSG_NOSIGNAL instead of using write(). puffs_framev_cmpframe_fn(pu, pufbuf_cmp1, pufbuf_cmp2, notresp) Compare the file system internal request tags in pufbuf_cmp1 and pufbuf_cmp2. Should return 0 if the tags are equal, 1 if first buffer's tag is greater than the second and -1 if it is smaller. The definitions "greater" and "smaller" are used transparently by the library, e.g. like qsort(3). If it can be determined from pufbuf_cmp1 that it is not a response to any outstanding request, notresp should be set to non-zero. This will cause puffs_framebuf to skip the test of the buffer against the rest of the outstanding request. May not block. puffs_framev_gotframe_fn(pu, pufbuf) Called when no outstanding request matches an incoming frame. The ownership of pufbuf is transferred to the called function and must be destroyed once processing is over. May not block. puffs_framev_fdnotify_fn(pu, fd, what) Is called when the read or write side of the file descriptor fd is closed. It is called once for each side, the bitmask parameter what specified what is currently closed: PUFFS_FBIO_READ and PUFFS_FBIO_WRITE for read and write, respectively. puffs_framev_init(pu, rfb, wfb, cmpfb, gotfb, fdnotfn) Initializes the given callbacks to the system. They will be used when puffs_mainloop() is called. The framework provides the routines puffs_framev_removeonclose() and puffs_framev_unmountonclose(), which can be given as fdnotfn. The first one removes the file descriptor once both sides are closed while the second one unmounts the file system and exits the mainloop. puffs_framev_addfd(pu, fd, what) Add file descriptor fd to be handled by the framework. It is legal to add a file descriptor either before calling puffs_mainloop() or at time when running. The parameter what controls enabling of input and output events and can be a bitwise combination of PUFFS_FBIO_READ and PUFFS_FBIO_WRITE. If not specified, the descriptor will be in a disabled state. puffs_framev_enablefd(pu, fd, what) Enable events of type what for file descriptor fd. puffs_framev_disablefd(pu, fd, what) Disable events of type what for file descriptor fd. puffs_framev_removefd(pu, fd, error) Remove file descriptor fd from the list of descriptors handled by the framework. Removing a file descriptor causes all operations blocked either on output or input to be released with the error value error. In case 0 is supplied as this parameter, ECONNRESET is used. The file system must explicitly remove each fd it has added. A good place to do this is puffs_framev_fdnotify_fn() or puffs_node_reclaim(), depending a little on the structure of the file system. puffs_framev_unmountonclose(pu, fd, what) This is library provided convenience routine for puffs_framev_fdnotify_fn(). It unmounts the file system when both the read and write side are closed. It is useful for file systems such as mount_psshfs(8) which depend on a single connection. RETURN VALUES These functions generally return -1 to signal error and set errno to indicate the type of error. CODE REFERENCES The current users of puffs_framebuf in the tree are mount_psshfs(8) and mount_9p(8). See src/usr.sbin/puffs/mount_psshfs and src/usr.sbin/puffs/mount_9p for the respective usage examples. SEE ALSO puffs(3), puffs_cc(3), puffs_ops(3) Antti Kantee, Using puffs for Implementing Client-Server Distributed File Systems, Helsinki University of Technology, Tech Report TKK-TKO-B157, September 2007. Antti Kantee, "Send and Receive of File System Protocols: Userspace Approach With puffs", Proceedings of AsiaBSDCon 2008, pp. 55-70, March 2008. NetBSD 8.99.34 September 6, 2008 NetBSD 8.99.34