PUFFS(3) Library Functions Manual PUFFS(3)


puffsPass-to-Userspace Framework File System development interface


puffs Convenience Library (libpuffs, -lpuffs)


#include <puffs.h>
struct puffs_usermount *
puffs_init(struct puffs_ops *pops, const char *mntfromname, const char *puffsname, void *private, uint32_t flags);
puffs_mount(struct puffs_usermount *pu, const char *dir, int mntflags, puffs_cookie_t root_cookie);
puffs_getselectable(struct puffs_usermount *pu);
puffs_setblockingmode(struct puffs_usermount *pu, int mode);
puffs_getstate(struct puffs_usermount *pu);
puffs_setstacksize(struct puffs_usermount *pu, size_t stacksize);
puffs_setroot(struct puffs_usermount *pu, struct puffs_node *node);
puffs_setrootinfo(struct puffs_usermount *pu, enum vtype vt, vsize_t vsize, dev_t rdev);
struct puffs_node *
puffs_getroot(struct puffs_usermount *pu);
void *
puffs_getspecific(struct puffs_usermount *pu);
puffs_setspecific(struct puffs_usermount *pu, void *private);
puffs_setmaxreqlen(struct puffs_usermount *pu, size_t maxreqlen);
puffs_getmaxreqlen(struct puffs_usermount *pu);
puffs_setfhsize(struct puffs_usermount *pu, size_t fhsize, int flags);
puffs_setncookiehash(struct puffs_usermount *pu, int nhashes);
puffs_ml_loop_fn(struct puffs_usermount *pu);
puffs_ml_setloopfn(struct puffs_usermount *pu, puffs_ml_loop_fn lfn);
puffs_ml_settimeout(struct puffs_usermount *pu, struct timespec *ts);
puffs_daemon(struct puffs_usermount *pu, int nochdir, int noclose);
puffs_mainloop(struct puffs_usermount *pu);
puffs_unmountonsignal(int sig, bool ignoresig);
puffs_dispatch_create(struct puffs_usermount *pu, struct puffs_framebuf *pb, struct puffs_cc **pccp);
puffs_dispatch_exec(struct puffs_cc *pcc, struct puffs_framebuf **pbp);


puffs provides a framework for creating file systems as userspace servers. Operations are transported from the kernel virtual file system layer to the concrete implementation behind puffs, where they are processed and results are sent back to the kernel.
It is possible to use puffs in two different ways. Calling puffs_mainloop() takes execution context away from the caller and automatically handles all requests by using the callbacks. By using puffs_framebuf(3) in conjuction with puffs_mainloop(), it is possible to handle I/O to and from file descriptors. This is suited e.g. for distributed file servers.

Library operation

Operations on the library always require a pointer to the opaque context identifier, struct puffs_usermount. It is obtained by calling puffs_init().
puffs operates using operation callbacks. They can be initialized using the macro PUFFSOP_SET(pops, fsname, type, opname), which will initialize the operation puffs_type_opname() in pops to fsname_type_opname(). All operations are initialized to a default state with the call PUFFSOP_INIT(pops). All of the VFS routines are mandatory, but all of the node operations with the exception of puffs_node_lookup() are optional. However, leaving operations blank will naturally have an effect on the features available from the file system implementation.
puffs_init(pops, mntfromname, puffsname, private, flags)
Initializes the library context. pops specifies the callback operations vector. mntfromname is device the file system is mounted from. This can be for example a block device such as /dev/wd0a or, if the file system is pseudo file system, the puffs device name can be given by _PATH_PUFFS. This value is used for example in the first column of the output of mount(8) and df(1). puffsname is the file system type. It will always be prepended with the string "puffs|". If possible, file server binaries should be named using the format "mount_myfsnamehere" and this value should equal "myfsnamehere". A file system specific context pointer can optionally be given in private. This can be retrieved by puffs_getspecific(). Flags for puffs can be given via flags. Currently the following flags are supported:
Do not enter pathname components into the name cache. This means that every time the kernel does a lookup for a componentname, the file server will be consulted.
Do not use the page cache. This means that all reads and writes to regular file are propagated to the file server for handling. This option makes a difference only for regular files.
This flag requests that all operations are sent to userspace. Normally the kernel shortcircuits unimplemented operations. This flag is mostly useful for debugging purposes.
Set the file system cache behavior as write-through. This means that all writes are immediately issued to the file server instead of being flushed in file system sync. This is useful especially for distributed file systems.
Issue inactive only on demand. If a file server defines the inactive method, call it only if the file server has explicitly requested that inactive be called for the node in question. Once inactive has been called for a node, it will not be called again unless the request to call inactive is reissued by the file server. See puffs_setback() in puffs_ops(3) for more information.
This flag affects only the parameter pcn to puffs_node_lookup(). If this flag is not given, only the next pathname component under lookup is found from pcn->pcn_name. If this flag is given, the full path the kernel was asked to resolve can be found from there.
The framework will build a complete path name, which is supplied with each operation and can be found from the pcn_po_full.po_path field in a struct puffs_cn. The option assumes that the framework can map a cookie to a struct puffs_node. See Cookies for more information on cookie mapping. See puffs_path(3) for more information on library calls involving paths.
Calculate a hash of the path into the path object field po_hash. This hash value is used by puffs_path_walkcmp() to avoid doing a full comparison for every path equal in length to the one searched for. Especially if the file system uses the abovementioned function, it is a good idea to define this flag.
Tell puffs that cookies map to struct pnode. This is automagically set if puffs_pn_new() is called.
Enforce name and attribute caches based on file system-supplied TTL. In lookup, create, mknod, mkdir, and symlink, the file system must update the node attributes, their TTL, and the node name TTL through puffs_newinfo_setva(), puffs_newinfo_setvattl(), and puffs_newinfo_setcnttl().
Additionally, puffs_node_getattr_ttl() and puffs_node_setattr_ttl() will be called instead of puffs_node_getattr() and puffs_node_setattr().
Never send lookups for “..” to the file system. Parent vnodes are all kept active until their children are reclaimed.
Do not send metadata cache flushes for time and size to the file system, which should take care of updating the values on its own.
This option makes the framework dump a textual representation of each operation before executing it. It is useful for debugging purposes.
The following functions can be used to query or modify the global state of the file system. Note, that all calls are not available at all times.
Returns a handle to do I/O multiplexing with: select(2), poll(2), and kqueue(2) are all examples of acceptable operations.
puffs_setblockingmode(pu, mode)
Sets the file system upstream access to blocking or non-blocking mode. Acceptable values for the argument are PUFFSDEV_BLOCK and PUFFSDEV_NONBLOCK.
This routine can be called only after calling puffs_mount().
Returns the state of the file system. It is maintained by the framework and is mostly useful for the framework itself. Possible values are PUFFS_STATE_BEFOREMOUNT, PUFFS_STATE_RUNNING, PUFFS_STATE_UNMOUNTING and PUFFS_STATE_UNMOUNTED.
puffs_setstacksize(pu, stacksize)
Sets the stack size used when running callbacks. The default is PUFFS_STACKSIZE_DEFAULT bytes of stack space per request. The minimum stacksize is architecture-dependent and can be specified by using the opaque constant PUFFS_STACKSIZE_MIN.
puffs_setroot(pu, node)
Sets the root node of mount pu to node. Setting the root node is currently required only if the path framework is used, see puffs_path(3).
puffs_setrootinfo(pu, vt, vsize, rdev)
The default root node is a directory. In case the file system wants something different, it can call this function and set the type, size and possible device type to whatever it wants. This routine is independent of puffs_setroot().
Returns the root node set earlier.
Returns the private argument of puffs_init().
puffs_setspecific(pu, private)
Can be used to set the specific data after the call to puffs_init().
puffs_setmaxreqlen(pu, maxreqlen)
In case the file system desires a maximum buffer length different from the default, the amount maxreqlen will be requested from the kernel when the file system is mounted.
It is legal to call this function only between puffs_init() and puffs_mount().
NOTE This does not currently work.
Returns the maximum request length the kernel will need for a single request.
NOTE This does not currently work.
puffs_setfhsize(pu, fhsize, flags)
Sets the desired file handle size. This must be called if the file system wishes to support NFS exporting file systems of the fh*() family of function calls.
In case all nodes in the file system produce the same length file handle, it must be supplied as fhsize. In this case, the file system may ignore the length parameters in the file handle callback routines, as the kernel will always pass the correct length buffer. However, if the file handle size varies according to file, the argument fhsize defines the maximum size of a file handle for the file system. In this case the file system must take care of the handle lengths by itself in the file handle callbacks, see puffs_ops(3) for more information. Also, the flag PUFFS_FHFLAG_DYNAMIC must be provided in the argument flags.
In case the file system wants to sanity check its file handle lengths for the limits of NFS, it can supply PUFFS_FHFLAG_NFSV2 and PUFFS_FHFLAG_NFSV3 in the flags parameter. It is especially important to note that these are not directly the limits specified by the protocols, as the kernel uses some bytes from the buffer space. In case the file handles are too large, mount will return an error.
It is legal to call this function only between puffs_init() and puffs_mount().
puffs_setncookiehash(pu, ncookiehash)
The parameter ncookiehash controls the amount of hash buckets the kernel has for reverse lookups from cookie to vnode. Technically the default is enough, but a memory/time tradeoff can be made by increasing this for file systems which know they will have very many active files.
It is legal to call this function only between puffs_init() and puffs_mount().
After the correct setup for the library has been established and the backend has been initialized the file system is made operational by calling puffs_mount(). After this function returns the file system should start processing requests.
puffs_mount(pu, dir, mntflags, root_cookie)
pu is the library context pointer from puffs_init(). The argument dir signifies the mount point and mntflags is the flagset given to mount(2). The value root_cookie will be used as the cookie for the file system root node.

Using the built-in eventloop

Loop function signature.
puffs_ml_setloopfn(pu, lfn)
Set loop function to lfn. This function is called once each time the event loop loops. It is not a well-defined interval, but it can be made fairly regular by setting the loop timeout by puffs_ml_settimeout().
puffs_ml_settimeout(pu, ts)
Sets the loop timeout to ts or disables it if ts is NULL. This can be used to roughly control how often the loop callback lfn() is called
puffs_daemon(pu, nochdir, noclose)
Detach from the console like daemon(3). This call synchronizes with puffs_mount() and the foreground process does not exit before the file system mount call has returned from the kernel. Since this routine internally calls fork, it has to be called before puffs_mount().
puffs_mainloop(pu, flags)
Handle all requests automatically until the file system is unmounted. It returns 0 if the file system was successfully unmounted or -1 if it was killed in action.
In case puffs_framebuf(3) has been initialized, I/O from the relevant descriptors is processed automatically by the eventloop.
puffs_unmountonsignal(signum, ignoresig)
Cause all file servers within the process to initiate unmount upon receipt of signal signum. This works only for servers which call puffs_mainloop() and must be called before any server within the process enters the mainloop. The process signal handler is still called before starting the unmount procedure. The parameter ignoresig is provided as a convenience and tells if to install a signal handler to ignore sig so that the process will not e.g. terminate based on the default action before the file system unmount can be initiated.
puffs_dispatch_create(pu, pb, pccp)
puffs_dispatch_exec(pcc, pbp)
In case the use of puffs_mainloop() is not possible, requests may be dispatched manually. However, as this is less efficient than using the mainloop, it should never be the first preference.
Calling puffs_dispatch_create() creates a dispatch request. The argument pb should contains a valid request and upon success pccp will contain a valid request context. This context is passed to puffs_dispatch_exec() to execute the request. If the request yielded before completing, the routine returns 0, otherwise 1. When the routine completes, pcc is made invalid and a pointer to the processed buffer is placed in pbp. It is the responsibility of the caller to send the response (if necessary) and destroy the buffer.
See puffs_cc(3) and puffs_framebuf(3) for further information.


Every file (regular file, directory, device node, ...) instance is attached to the kernel using a cookie. A cookie should uniquely map to a file during its lifetime. If file instances are kept in memory, a simple strategy is to use the virtual address of the structure describing the file. The cookie can be recycled when puffs_node_reclaim() is called for a node.
For some operations (such as building paths) the framework needs to map the cookie to the framework-level structure describing a file, struct puffs_node. It is advisable to simply use the struct puffs_node address as a cookie and store file system specific data in the private portion of struct puffs_node. The library assumes this by default. If it is not desirable, the file system implementation can call puffs_set_cookiemap() to provide an alternative cookie-to-node mapping function.


mount(2), puffs_cc(3), puffs_cred(3), puffs_flush(3), puffs_framebuf(3), puffs_node(3), puffs_ops(3), puffs_path(3), refuse(3), puffs(4)
Antti Kantee, puffs - Pass-to-Userspace Framework File System, Proceedings of AsiaBSDCon 2007, pp. 29-42, March 2007.
Antti Kantee, Using puffs for Implementing Client-Server Distributed File Systems, Helsinki University of Technology, Tech Report TKK-TKO-B157, September 2007.
Antti Kantee and Alistair Crooks, ReFUSE: Userspace FUSE Reimplementation Using puffs, EuroBSDCon 2007, September 2007.
Antti Kantee, Send and Receive of File System Protocols: Userspace Approach With puffs, Proceedings of AsiaBSDCon 2008, pp. 55-70, March 2008.


An unsupported experimental version of puffs first appeared in NetBSD 4.0. A stable version appeared in NetBSD 5.0.


Antti Kantee <pooka@iki.fi>
April 10, 2016 NetBSD-current