File system callbacks (puffs_fs)

puffs_fs_statvfs(pu, sbp)

The following fields of the argument sbp need to be filled:

 * unsigned long  f_bsize;     file system block size
 * unsigned long  f_frsize;    fundamental file system block size
 * fsblkcnt_t     f_blocks;    number of blocks in file system,
 *                                      (in units of f_frsize)
 *
 * fsblkcnt_t     f_bfree;     free blocks avail in file system
 * fsblkcnt_t     f_bavail;    free blocks avail to non-root
 * fsblkcnt_t     f_bresvd;    blocks reserved for root

 * fsfilcnt_t     f_files;     total file nodes in file system
 * fsfilcnt_t     f_ffree;     free file nodes in file system
 * fsfilcnt_t     f_favail;    free file nodes avail to non-root
 * fsfilcnt_t     f_fresvd;    file nodes reserved for root

    

puffs_fs_sync(pu, waitfor, pcr)

All the dirty buffers that have been cached at the file server level including metadata should be committed to stable storage. The waitfor parameter affects the operation. Possible values are:

MNT_WAIT

Wait for all I/O for complete until returning.

MNT_NOWAIT

Initiate I/O, but do not wait for completion.

MNT_LAZY

Synchronize data not synchronized by the file system syncer, i.e., data not written when node_fsync() is called with FSYNC_LAZY.

The credentials for the initiator of the sync operation are present in pcr and will usually be either file system or kernel credentials, but might also be user credentials. However, most of the time it is advisable to sync regardless of the credentials of the caller.

puffs_fs_fhtonode(pu, fid, fidsize, pni)

Translates a file handle fid to a node. The parameter fidsize indicates how large the file handle is. In case the file system's handles are static length, this parameter can be ignored as the kernel guarantees all file handles passed to the file server are of correct length. For dynamic length handles the field should be examined and EINVAL returned in case the file handle length is not correct.

This function provides essentially the same information to the kernel as puffs_node_lookup(). The information is necessary for creating a new vnode corresponding to the file handle.

puffs_fs_nodetofh(pu, cookie, fid, fidsize)

Create a file handle from the node described by cookie. The file handle should contain enough information to reliably identify the node even after reboots and the pathname/inode being replaced by another file. If this is not possible, it is up to the author of the file system to act responsibly and decide if the file system can support file handles at all.

For file systems which want dynamic length file handles, this function must check if the file handle space indicated by fidsize is large enough to accommodate the file handle for the node. If not, it must fill in the correct size and return E2BIG. In either case, the handle length should be supplied to the kernel in fidsize. File systems with static length handles can ignore the size parameter as the kernel always supplies the correct size buffer.

puffs_fs_unmount(pu, flags)

Unmount the file system. The kernel has assumedly flushed all cached data when this callback is executed. If the file system cannot currently be safely be unmounted, for whatever reason, the kernel will honor an error value and not forcibly unmount. However, if the flag MNT_FORCE is not honored by the file server, the kernel will forcibly unmount the file system.

Node callbacks

These operations operate in the level of individual files. The file cookie is always provided as the second argument opc. If the operation is for a file, it will be the cookie of the file. The case the operation involves a directory (such as “create file in directory”), the cookie will be for the directory. Some operations take additional cookies to describe the rest of the operands. The return value 0 signals success, else an appropriate errno value should be returned. Please note that neither this list nor the descriptions are complete.

puffs_node_lookup(pu, opc, pni, pcn)

This function is used to locate nodes, or in other words translate pathname components to file system data structures. The implementation should match the name in pcn against the existing entries in the directory provided by the cookie opc. If found, the cookie for the located node should be set in pni using puffs_newinfo_setcookie(). Additionally, the vnode type and size (latter applicable to regular files only) should be set using puffs_newinfo_setvtype() and puffs_newinfo_setsize(), respectively. If the located entry is a block device or character device file, the dev_t for the entry should be set using puffs_newinfo_setrdev().

If puffs_init() was called with PUFFS_KFLAG_CACHE_FS_TTL then puffs_newinfo_setva(), puffs_newinfo_setvattl(), and puffs_newinfo_setcnttl() can be called to specify the new node attributes, cached attributes time to live, and cached name time to live.

The type of operation is found from pcn->pcn_nameiop:

NAMEI_LOOKUP

Normal lookup operation.

NAMEI_CREATE

Lookup to create a node.

NAMEI_DELETE

Lookup for node deletion.

NAMEI_RENAME

Lookup for the target of a rename operation (source will be looked up using NAMEI_DELETE).

The final component from a pathname lookup usually requires special treatment. It can be identified by looking at the pcn->pcn_flags fields for the flag PUFFSLOOKUP_ISLASTCN. For example, in most cases the lookup operation will want to check if a delete, rename or create operation has enough credentials to perform the operation.

The return value 0 signals a found node and a nonzero value signals an errno. As a special case, ENOENT signals "success" for cases where the lookup operation is NAMEI_CREATE or NAMEI_RENAME. Failure in these cases can be signalled by returning another appropriate error code, for example EACCESS.

Usually a null-terminated string for the next pathname component is provided in pcn->pcn_name. In case the file system is using the option PUFFS_KFLAG_LOOKUP_FULLPNBUF, the remainder of the complete pathname under lookup is found in the same location. pcn->pcn_namelen always specifies the length of the next component. If operating with a full path, the file system is allowed to consume more than the next component's length in node lookup. This is done by setting pcn->pcn_consume to indicate the amount of extra characters in addition to pcn->pcn_namelen processed.

puffs_node_create(pu, opc, pni, pcn, va)

 

puffs_node_mkdir(pu, opc, pni, pcn, va)

 

puffs_node_mknod(pu, opc, pni, pcn, va)

A file node is created in the directory denoted by the cookie opc by any of the above callbacks. The name of the new file can be found from pcn and the attributes are specified by va and the cookie for the newly created node should be set in pni. The only difference between these three is that they create a regular file, directory and device special file, respectively.

In case of mknod, the device identifier can be found in va->va_rdev.

puffs_node_open(pu, opc, mode, pcr)

 

puffs_node_open2(pu, opc, mode, pcr, oflags)

Open the node denoted by the cookie opc. The parameter mode specifies the flags that open(2) was called with, e.g. O_APPEND and O_NONBLOCK. puffs_node_open2() allows the file system to pass back information for the file in oflags. The only implemented flag for now is PUFFS_OPEN_IO_DIRECT that cause file read/write to bypass the page cache.

puffs_node_close(pu, opc, flags, pcr)

Close a node. The parameter flags parameter describes the flags that the file was opened with.

puffs_node_access(pu, opc, mode, pcr)

Check if the credentials of pcr have the right to perform the operation specified by mode onto the node opc. The argument mode can specify read, write or execute by PUFFS_VREAD, PUFFS_VWRITE, and PUFFS_VEXEC, respectively.

puffs_node_getattr(pu, opc, va, pcr)

The attributes of the node specified by opc must be copied to the space pointed by va.

puffs_node_getattr_ttl(pu, opc, va, pcr, va_ttl)

Same as puffs_node_getattr() with cached attribute time to live specified in va_ttl

puffs_node_setattr(pu, opc, va, pcr)

The attributes for the node specified by opc must be set to those contained in va. Only fields of va which contain a value different from PUFFS_VNOVAL (typecast to the field's type!) contain a valid value.

puffs_node_setattr_ttl(pu, opc, va, pcr, va_ttl, xflag)

Same as puffs_node_setattr() with cached attribute time to live specified in va_ttl. PUFFS_SETATTR_FAF will be set in xflag for Fire-And-Forget operations.

puffs_node_pathconf(pu, opc, name, retval)

The value of the pathconf(2) filesystem limit specified in name for the node opc should be copied to the space pointed to by retval.

puffs_node_advlock(po, opc, id, op, fl, flags)

Manipulate advisory record locks on the node opc. The argument id is the id token which is changing the lock, op is the fcntl(2) operation to perform, fl is the lock descriptor structure and flags are the flags as defined in VOP_ADVLOCK(9).

puffs_node_poll(pu, opc, events)

Poll for events on node opc. If poll(2) events specified in events are available, the function should set the bitmask to match available events and return immediately. Otherwise, the function should block (yield) until some events in events become available and only then set the events bitmask and return.

In case this function returns an error, POLLERR (or its select(2) equivalent) will be delivered to the calling process.

NOTE! The system call interface for poll() contains a timeout parameter. At this level, however, the timeout is not supplied. In case input does not arrive, the file system should periodically unblock and return 0 new events to avoid hanging forever. This will hopefully be better supported by libpuffs in the future.

puffs_node_mmap(pu, opc, flags, pcr)

Called when a regular file is being memory mapped by mmap(2). flags is currently always 0.

puffs_node_fsync(pu, opc, pcr, flags, offlo, offhi)

Synchronize a node's contents onto stable storage. This is necessary only if the file server caches some information before committing it. The parameter flags specifies the minimum level of synchronization required (XXX: they are not yet available). The parameters offlo and offhi specify the data offsets requiring to be synced. A high offset of 0 means sync from offlo to the end of the file.

puffs_node_seek(pu, opc, oldoff, newoff, pcr)

Test if the node opc is seekable to the location newoff. The argument oldoff specifies the offset we are starting the seek from. Most file systems dealing only with regular will choose to not implement this. However, it is useful for example in cases where files are unseekable streams.

puffs_node_remove(pu, opc, targ, pcn)

 

puffs_node_rmdir(pu, opc, targ, pcn)

Remove the node targ from the directory indicated by opc. The directory entry name to be removed is provided by pcn. The rmdir operation removes only directories, while the remove operation removes all other types except directories.

It is paramount to note that the file system may not remove the node data structures at this point, only the directory entry and prevent lookups from finding the node again. This is to retain the UNIX open file semantics. The data may be removed only when puffs_node_reclaim() or puffs_node_reclaim2() is called for the node, as this assures there are no further users.

puffs_node_link(pu, opc, targ, pcn)

Create a hard link for the node targ into the directory opc. The argument pcn provides the directory entry name for the new link.

puffs_node_rename(pu, src_dir, src, pcn_src, targ_dir, targ, pcn_targ)

Rename the node src with the name specified by pcn_src from the directory src_dir. The target directory and target name are given by targ_dir and pcn_targ, respectively. If the target node already exists, it is specified by targ and must be replaced atomically. Otherwise targ is gives as NULL.

It is legal to replace a directory node by another directory node with the means of rename if the target directory is empty, otherwise ENOTEMPTY should be returned. All other types can replace all other types. In case a rename between incompatible types is attempted, the errors ENOTDIR or EISDIR should be returned, depending on the target type.

puffs_node_readdir(pu, opc, dent, readoff, reslen, pcr, eofflag, cookies, ncookies)

To read directory entries, puffs_node_readdir() is called. It should store directories as struct dirent in the space pointed to by dent. The amount of space available is given by reslen and before returning it should be set to the amount of space remaining in the buffer. The argument offset is used to specify the offset to the directory. Its interpretation is up to the file system and it should be set to signal the continuation point when there is no more room for the next entry in dent. It is most performant to return the maximal amount of directory entries each call. It is easiest to generate directory entries by using puffs_nextdent(), which also automatically advances the necessary pointers.

In case end-of-directory is reached, eofflag should be set to one. Note that even a new call to readdir may start where readoff points to end-of-directory.

If the file system supports file handles, the arguments cookies and ncookies must be filled out. cookies is a vector for offsets corresponding to read offsets. One cookie should be filled out for each directory entry. The value of the cookie should equal the offset of the next directory entry, i.e., which offset should be passed to readdir for the first entry read to be the entry following the current one. ncookies is the number of slots for cookies in the cookie vector upon entry to the function and must be set to the amount of cookies stored in the vector (i.e., amount of directory entries read) upon exit. There is always enough space in the cookie vector for the maximal number of entries that will fit into the directory entry buffer. For filling out the vector, the helper function PUFFS_STORE_DCOOKIE(cookies, ncookies, offset) can be used. This properly checks against cookies being NULL. Note that ncookies must be initialized to zero before the first call to PUFFS_STORE_DCOOKIE().

puffs_node_symlink(pu, opc, pni, pcn_src, va, link_target)

Create a symbolic link into the directory opc with the name in pcn_src and the initial attributes in va. The argument link_target contains a null-terminated string for the link target. The created node cookie should be set in pni.

puffs_node_readlink(pu, opc, pcr, link, linklen)

Read the target of a symbolic link opc. The result is placed in the buffer pointed to by link. This buffer's length is given in linklen and it must be updated to reflect the real link length. A terminating nul character should not be put into the buffer and must not be included in the link length.

puffs_node_read(pu, opc, buf, offset, resid, pcr, ioflag)

Read the contents of a file opc. It will gather the data from offset in the file and read the number resid octets. The buffer is guaranteed to have this much space. The amount of data requested by resid should be read, except in the case of eof-of-file or an error. The parameter resid should be set to indicate the amount of request NOT completed. In the normal case this should be 0.

puffs_node_write(pu, opc, buf, offset, resid, pcr, ioflag)

 

puffs_node_write2(pu, opc, buf, offset, resid, pcr, ioflag, xflag)

puffs_node_write() writes data to a file opc at offset and extend the file if necessary. The number of octets written is indicated by resid; everything must be written or an error will be generated. The parameter must be set to indicate the amount of data NOT written. In case the ioflag PUFFS_IO_APPEND is specified, the data should be appended to the end of the file. puffs_node_write2() serves the same purpose as puffs_node_write() with an additional xflag in which PUFFS_WRITE_FAF is set for Fire-And-Forget operations.

puffs_node_fallocate(pu, opc, pos, len)

Allocate len bytes of backing store at offset pos for the node referenced by the cookie opc.

puffs_node_fdiscard(pu, opc, pos, len)

Free len bytes of backing store (creating a hole) at offset pos for the node referenced by the cookie opc.

puffs_node_print(pu, opc)

Print information about node. This is used only for kernel-initiated diagnostic purposes.

puffs_node_reclaim(pu, opc)

The kernel will no longer reference the cookie and resources associated with it may be freed. In case the file opc has a link count of zero, it may be safely removed now.

puffs_node_reclaim2(pu, opc, nlookup)

Same as puffs_node_reclaim() with an additional argument for the number of lookups that have been done on the node (Node creation is counted as a lookup). This can be used by the file system to avoid a race condition, where the kernel sends a reclaim while it does not have received the reply for a lookup. If the file system tracks lookup count, and compares to nlookup it can detect this situation and ignore the reclaim.

If the file system maps cookies to struct puffs_node then the framework will do that work, and puffs_node_reclaim() can be reliably used without the race condition.

puffs_node_abortop(pu, opc, pcn)

In case the operation following lookup (e.g., mkdir or remove) is not executed for some reason, abortop will be issued. This is useful only for servers which cache state between lookup and a directory operation and is generally left unimplemented.

puffs_node_inactive(pu, opc)

The node opc has lost its last reference in the kernel. However, the cookie must still remain valid until puffs_node_reclaim() or puffs_node_reclaim2() is called.

puffs_setback(pcc, op)

Issue a "setback" operation which will be handled when the request response is returned to the kernel. Currently this can be only called from mmap, open, remove and rmdir. The valid parameters for op are PUFFS_SETBACK_INACT_N1 and PUFFS_SETBACK_INACT_N2. These signal that a file system mounted with PUFFS_KFLAG_IAONDEMAND should call the file system inactive method for the specified node. The node number 1 always means the operation cookie opc, while the node number 2 can be used to specify the second node argument present in some methods, e.g., remove.

puffs_newinfo_setcookie(pni, cookie)

Set cookie for node provided by this method to cookie.

puffs_newinfo_setvtype(pni, vtype)

Set the type of the newly located node to vtype. This call is valid only for lookup() and fhtonode().

puffs_newinfo_setsize(pni, size)

Set the size of the newly located node to size. If left unset, the value defaults to 0. This call is valid only for lookup() and fhtovp().

puffs_newinfo_setrdev(pni, rdev)

Set the type of the newly located node to vtype. This call is valid only for lookup() and fhtovp() producing device type nodes.

puffs_newinfo_setva(pni, vap)

Set the attributes for newly created vnode. This call is valid for lookup(), create(), mkdir(), mknod(), and symlink(), if puffs_init() was called with PUFFS_KFLAG_CACHE_FS_TTL flag set.

puffs_newinfo_setvattl(pni, va_ttl)

Set cached attribute time to live for newly created vnode. This call is valid for lookup(), create(), mkdir(), mknod(), and symlink(), if puffs_init() was called with PUFFS_KFLAG_CACHE_FS_TTL flag set.

puffs_newinfo_setcnttl(pni, cn_ttl)

Set cached name time to live for newly created vnode. This call is valid for lookup(), create(), mkdir(), mknod(), and symlink(), if puffs_init() was called with PUFFS_KFLAG_CACHE_FS_TTL flag set.