M_SPLIT(9FREEBSD) - man page online | system kernel interfaces
Memory management in the kernel IPC subsystem.
November 13, 2012
MBUF(9) BSD Kernel Developer's Manual MBUF(9)
NAMEmbuf — memory management in the kernel IPC subsystem
SYNOPSIS#include <sys/param.h> #include <sys/systm.h> #include <sys/mbuf.h> Mbuf allocation macros MGET(struct mbuf *mbuf, int how, short type); MGETHDR(struct mbuf *mbuf, int how, short type); MCLGET(struct mbuf *mbuf, int how); MEXTADD(struct mbuf *mbuf, caddr_t buf, u_int size, void (*free)(void *opt_arg1, void *opt_arg2), void *opt_arg1, void *opt_arg2, short flags, int type); MEXTFREE(struct mbuf *mbuf); MFREE(struct mbuf *mbuf, struct mbuf *successor); Mbuf utility macros mtod(struct mbuf *mbuf, type); M_ALIGN(struct mbuf *mbuf, u_int len); MH_ALIGN(struct mbuf *mbuf, u_int len); int M_LEADINGSPACE(struct mbuf *mbuf); int M_TRAILINGSPACE(struct mbuf *mbuf); M_MOVE_PKTHDR(struct mbuf *to, struct mbuf *from); M_PREPEND(struct mbuf *mbuf, int len, int how); MCHTYPE(struct mbuf *mbuf, u_int type); int M_WRITABLE(struct mbuf *mbuf); Mbuf allocation functions struct mbuf * m_get(int how, int type); struct mbuf * m_getm(struct mbuf *orig, int len, int how, int type); struct mbuf * m_getcl(int how, short type, int flags); struct mbuf * m_getclr(int how, int type); struct mbuf * m_gethdr(int how, int type); struct mbuf * m_free(struct mbuf *mbuf); void m_freem(struct mbuf *mbuf); Mbuf utility functions void m_adj(struct mbuf *mbuf, int len); void m_align(struct mbuf *mbuf, int len); int m_append(struct mbuf *mbuf, int len, c_caddr_t cp); struct mbuf * m_prepend(struct mbuf *mbuf, int len, int how); struct mbuf * m_copyup(struct mbuf *mbuf, int len, int dstoff); struct mbuf * m_pullup(struct mbuf *mbuf, int len); struct mbuf * m_pulldown(struct mbuf *mbuf, int offset, int len, int *offsetp); struct mbuf * m_copym(struct mbuf *mbuf, int offset, int len, int how); struct mbuf * m_copypacket(struct mbuf *mbuf, int how); struct mbuf * m_dup(struct mbuf *mbuf, int how); void m_copydata(const struct mbuf *mbuf, int offset, int len, caddr_t buf); void m_copyback(struct mbuf *mbuf, int offset, int len, caddr_t buf); struct mbuf * m_devget(char *buf, int len, int offset, struct ifnet *ifp, void (*copy)(char *from, caddr_t to, u_int len)); void m_cat(struct mbuf *m, struct mbuf *n); u_int m_fixhdr(struct mbuf *mbuf); void m_dup_pkthdr(struct mbuf *to, struct mbuf *from); void m_move_pkthdr(struct mbuf *to, struct mbuf *from); u_int m_length(struct mbuf *mbuf, struct mbuf **last); struct mbuf * m_split(struct mbuf *mbuf, int len, int how); int m_apply(struct mbuf *mbuf, int off, int len, int (*f)(void *arg, void *data, u_int len), void *arg); struct mbuf * m_getptr(struct mbuf *mbuf, int loc, int *off); struct mbuf * m_defrag(struct mbuf *m0, int how); struct mbuf * m_unshare(struct mbuf *m0, int how);
DESCRIPTIONAn mbuf is a basic unit of memory management in the kernel IPC subsystem. Network packets and socket buffers are stored in mbufs. A network packet may span multiple mbufs arranged into a mbuf chain (linked list), which allows adding or trimming network headers with little overhead. While a developer should not bother with mbuf internals without serious reason in order to avoid incompatibilities with future changes, it is useful to understand the general struc‐ ture of an mbuf. An mbuf consists of a variable-sized header and a small internal buffer for data. The total size of an mbuf, MSIZE, is a constant defined in <sys/param.h>. The mbuf header includes: m_next (struct mbuf *) A pointer to the next mbuf in the mbuf chain. m_nextpkt (struct mbuf *) A pointer to the next mbuf chain in the queue. m_data (caddr_t) A pointer to data attached to this mbuf. m_len (int) The length of the data. m_type (short) The type of the data. m_flags (int) The mbuf flags. The mbuf flag bits are defined as follows: /* mbuf flags */ #define M_EXT 0x0001 /* has associated external storage */ #define M_PKTHDR 0x0002 /* start of record */ #define M_EOR 0x0004 /* end of record */ #define M_RDONLY 0x0008 /* associated data marked read-only */ #define M_PROTO1 0x0010 /* protocol-specific */ #define M_PROTO2 0x0020 /* protocol-specific */ #define M_PROTO3 0x0040 /* protocol-specific */ #define M_PROTO4 0x0080 /* protocol-specific */ #define M_PROTO5 0x0100 /* protocol-specific */ #define M_PROTO6 0x4000 /* protocol-specific (avoid M_BCAST conflict) */ #define M_FREELIST 0x8000 /* mbuf is on the free list */ /* mbuf pkthdr flags (also stored in m_flags) */ #define M_BCAST 0x0200 /* send/received as link-level broadcast */ #define M_MCAST 0x0400 /* send/received as link-level multicast */ #define M_FRAG 0x0800 /* packet is fragment of larger packet */ #define M_FIRSTFRAG 0x1000 /* packet is first fragment */ #define M_LASTFRAG 0x2000 /* packet is last fragment */ The available mbuf types are defined as follows: /* mbuf types */ #define MT_DATA 1 /* dynamic (data) allocation */ #define MT_HEADER MT_DATA /* packet header */ #define MT_SONAME 8 /* socket name */ #define MT_CONTROL 14 /* extra-data protocol message */ #define MT_OOBDATA 15 /* expedited data */ The available external buffer types are defined as follows: /* external buffer types */ #define EXT_CLUSTER 1 /* mbuf cluster */ #define EXT_SFBUF 2 /* sendfile(2)'s sf_bufs */ #define EXT_JUMBOP 3 /* jumbo cluster 4096 bytes */ #define EXT_JUMBO9 4 /* jumbo cluster 9216 bytes */ #define EXT_JUMBO16 5 /* jumbo cluster 16184 bytes */ #define EXT_PACKET 6 /* mbuf+cluster from packet zone */ #define EXT_MBUF 7 /* external mbuf reference (M_IOVEC) */ #define EXT_NET_DRV 100 /* custom ext_buf provided by net driver(s) */ #define EXT_MOD_TYPE 200 /* custom module's ext_buf type */ #define EXT_DISPOSABLE 300 /* can throw this buffer away w/page flipping */ #define EXT_EXTREF 400 /* has externally maintained ref_cnt ptr */ If the M_PKTHDR flag is set, a struct pkthdr m_pkthdr is added to the mbuf header. It con‐ tains a pointer to the interface the packet has been received from (struct ifnet *rcvif), and the total packet length (int len). Optionally, it may also contain an attached list of packet tags (struct m_tag). See mbuf_tags(9) for details. Fields used in offloading check‐ sum calculation to the hardware are kept in m_pkthdr as well. See HARDWARE-ASSISTED CHECKSUM CALCULATION for details. If small enough, data is stored in the internal data buffer of an mbuf. If the data is suf‐ ficiently large, another mbuf may be added to the mbuf chain, or external storage may be associated with the mbuf. MHLEN bytes of data can fit into an mbuf with the M_PKTHDR flag set, MLEN bytes can otherwise. If external storage is being associated with an mbuf, the m_ext header is added at the cost of losing the internal data buffer. It includes a pointer to external storage, the size of the storage, a pointer to a function used for freeing the storage, a pointer to an optional argument that can be passed to the function, and a pointer to a reference counter. An mbuf using external storage has the M_EXT flag set. The system supplies a macro for allocating the desired external storage buffer, MEXTADD. The allocation and management of the reference counter is handled by the subsystem. The system also supplies a default type of external storage buffer called an mbuf cluster. Mbuf clusters can be allocated and configured with the use of the MCLGET macro. Each mbuf cluster is MCLBYTES in size, where MCLBYTES is a machine-dependent constant. The system defines an advisory macro MINCLSIZE, which is the smallest amount of data to put into an mbuf cluster. It is equal to MHLEN plus one. It is typically preferable to store data into the data region of an mbuf, if size permits, as opposed to allocating a separate mbuf cluster to hold the same data. Macros and Functions There are numerous predefined macros and functions that provide the developer with common utilities. mtod(mbuf, type) Convert an mbuf pointer to a data pointer. The macro expands to the data pointer cast to the pointer of the specified type. Note: It is advisable to ensure that there is enough contiguous data in mbuf. See m_pullup() for details. MGET(mbuf, how, type) Allocate an mbuf and initialize it to contain internal data. mbuf will point to the allocated mbuf on success, or be set to NULL on failure. The how argument is to be set to M_WAITOK or M_NOWAIT. It specifies whether the caller is willing to block if necessary. A number of other functions and macros related to mbufs have the same argument because they may at some point need to allocate new mbufs. Historical mbuf allocator (See HISTORY section) used allocation flags M_WAIT and M_DONTWAIT. These constants are kept for compatibility and their use in new code is discouraged. MGETHDR(mbuf, how, type) Allocate an mbuf and initialize it to contain a packet header and internal data. See MGET() for details. MEXTADD(mbuf, buf, size, free, opt_arg1, opt_arg2, flags, type) Associate externally managed data with mbuf. Any internal data contained in the mbuf will be discarded, and the M_EXT flag will be set. The buf and size arguments are the address and length, respectively, of the data. The free argument points to a function which will be called to free the data when the mbuf is freed; it is only used if type is EXT_EXTREF. The opt_arg1 and opt_arg2 arguments will be passed unmodified to free. The flags argument specifies additional mbuf flags; it is not necessary to specify M_EXT. Finally, the type argument specifies the type of external data, which controls how it will be disposed of when the mbuf is freed. In most cases, the correct value is EXT_EXTREF. MCLGET(mbuf, how) Allocate and attach an mbuf cluster to mbuf. If the macro fails, the M_EXT flag will not be set in mbuf. M_ALIGN(mbuf, len) Set the pointer mbuf->m_data to place an object of the size len at the end of the internal data area of mbuf, long word aligned. Applicable only if mbuf is newly allo‐ cated with MGET() or m_get(). MH_ALIGN(mbuf, len) Serves the same purpose as M_ALIGN() does, but only for mbuf newly allocated with MGETHDR() or m_gethdr(), or initialized by m_dup_pkthdr() or m_move_pkthdr(). m_align(mbuf, len) Services the same purpose as M_ALIGN() but handles any type of mbuf. M_LEADINGSPACE(mbuf) Returns the number of bytes available before the beginning of data in mbuf. M_TRAILINGSPACE(mbuf) Returns the number of bytes available after the end of data in mbuf. M_PREPEND(mbuf, len, how) This macro operates on an mbuf chain. It is an optimized wrapper for m_prepend() that can make use of possible empty space before data (e.g. left after trimming of a link- layer header). The new mbuf chain pointer or NULL is in mbuf after the call. M_MOVE_PKTHDR(to, from) Using this macro is equivalent to calling m_move_pkthdr(to, from). M_WRITABLE(mbuf) This macro will evaluate true if mbuf is not marked M_RDONLY and if either mbuf does not contain external storage or, if it does, then if the reference count of the stor‐ age is not greater than 1. The M_RDONLY flag can be set in mbuf->m_flags. This can be achieved during setup of the external storage, by passing the M_RDONLY bit as a flags argument to the MEXTADD() macro, or can be directly set in individual mbufs. MCHTYPE(mbuf, type) Change the type of mbuf to type. This is a relatively expensive operation and should be avoided. The functions are: m_get(how, type) A function version of MGET() for non-critical paths. m_getm(orig, len, how, type) Allocate len bytes worth of mbufs and mbuf clusters if necessary and append the resulting allocated mbuf chain to the mbuf chain orig, if it is non-NULL. If the allocation fails at any point, free whatever was allocated and return NULL. If orig is non-NULL, it will not be freed. It is possible to use m_getm() to either append len bytes to an existing mbuf or mbuf chain (for example, one which may be sitting in a pre-allocated ring) or to simply perform an all-or-nothing mbuf and mbuf cluster allocation. m_gethdr(how, type) A function version of MGETHDR() for non-critical paths. m_getcl(how, type, flags) Fetch an mbuf with a mbuf cluster attached to it. If one of the allocations fails, the entire allocation fails. This routine is the preferred way of fetching both the mbuf and mbuf cluster together, as it avoids having to unlock/relock between alloca‐ tions. Returns NULL on failure. m_getclr(how, type) Allocate an mbuf and zero out the data region. m_free(mbuf) Frees mbuf. Returns m_next of the freed mbuf. The functions below operate on mbuf chains. m_freem(mbuf) Free an entire mbuf chain, including any external storage. m_adj(mbuf, len) Trim len bytes from the head of an mbuf chain if len is positive, from the tail other‐ wise. m_append(mbuf, len, cp) Append len bytes of data cp to the mbuf chain. Extend the mbuf chain if the new data does not fit in existing space. m_prepend(mbuf, len, how) Allocate a new mbuf and prepend it to the mbuf chain, handle M_PKTHDR properly. Note: It does not allocate any mbuf clusters, so len must be less than MLEN or MHLEN, depending on the M_PKTHDR flag setting. m_copyup(mbuf, len, dstoff) Similar to m_pullup() but copies len bytes of data into a new mbuf at dstoff bytes into the mbuf. The dstoff argument aligns the data and leaves room for a link layer header. Returns the new mbuf chain on success, and frees the mbuf chain and returns NULL on failure. Note: The function does not allocate mbuf clusters, so len + dstoff must be less than MHLEN. m_pullup(mbuf, len) Arrange that the first len bytes of an mbuf chain are contiguous and lay in the data area of mbuf, so they are accessible with mtod(mbuf, type). It is important to remem‐ ber that this may involve reallocating some mbufs and moving data so all pointers ref‐ erencing data within the old mbuf chain must be recalculated or made invalid. Return the new mbuf chain on success, NULL on failure (the mbuf chain is freed in this case). Note: It does not allocate any mbuf clusters, so len must be less than or equal to MHLEN. m_pulldown(mbuf, offset, len, offsetp) Arrange that len bytes between offset and offset + len in the mbuf chain are contigu‐ ous and lay in the data area of mbuf, so they are accessible with mtod(mbuf, type). len must be smaller than, or equal to, the size of an mbuf cluster. Return a pointer to an intermediate mbuf in the chain containing the requested region; the offset in the data region of the mbuf chain to the data contained in the returned mbuf is stored in *offsetp. If offp is NULL, the region may be accessed using mtod(mbuf, type). If offp is non-NULL, the region may be accessed using mtod(mbuf, uint8_t, +, *offsetp). The region of the mbuf chain between its beginning and off is not modified, therefore it is safe to hold pointers to data within this region before calling m_pulldown(). m_copym(mbuf, offset, len, how) Make a copy of an mbuf chain starting offset bytes from the beginning, continuing for len bytes. If len is M_COPYALL, copy to the end of the mbuf chain. Note: The copy is read-only, because the mbuf clusters are not copied, only their reference counts are incremented. m_copypacket(mbuf, how) Copy an entire packet including header, which must be present. This is an optimized version of the common case m_copym(mbuf, 0, M_COPYALL, how). Note: the copy is read- only, because the mbuf clusters are not copied, only their reference counts are incre‐ mented. m_dup(mbuf, how) Copy a packet header mbuf chain into a completely new mbuf chain, including copying any mbuf clusters. Use this instead of m_copypacket() when you need a writable copy of an mbuf chain. m_copydata(mbuf, offset, len, buf) Copy data from an mbuf chain starting off bytes from the beginning, continuing for len bytes, into the indicated buffer buf. m_copyback(mbuf, offset, len, buf) Copy len bytes from the buffer buf back into the indicated mbuf chain, starting at offset bytes from the beginning of the mbuf chain, extending the mbuf chain if neces‐ sary. Note: It does not allocate any mbuf clusters, just adds mbufs to the mbuf chain. It is safe to set offset beyond the current mbuf chain end: zeroed mbufs will be allocated to fill the space. m_length(mbuf, last) Return the length of the mbuf chain, and optionally a pointer to the last mbuf. m_dup_pkthdr(to, from, how) Upon the function's completion, the mbuf to will contain an identical copy of from->m_pkthdr and the per-packet attributes found in the mbuf chain from. The mbuf from must have the flag M_PKTHDR initially set, and to must be empty on entry. m_move_pkthdr(to, from) Move m_pkthdr and the per-packet attributes from the mbuf chain from to the mbuf to. The mbuf from must have the flag M_PKTHDR initially set, and to must be empty on entry. Upon the function's completion, from will have the flag M_PKTHDR and the per- packet attributes cleared. m_fixhdr(mbuf) Set the packet-header length to the length of the mbuf chain. m_devget(buf, len, offset, ifp, copy) Copy data from a device local memory pointed to by buf to an mbuf chain. The copy is done using a specified copy routine copy, or bcopy() if copy is NULL. m_cat(m, n) Concatenate n to m. Both mbuf chains must be of the same type. N is still valid after the function returned. Note: It does not handle M_PKTHDR and friends. m_split(mbuf, len, how) Partition an mbuf chain in two pieces, returning the tail: all but the first len bytes. In case of failure, it returns NULL and attempts to restore the mbuf chain to its original state. m_apply(mbuf, off, len, f, arg) Apply a function to an mbuf chain, at offset off, for length len bytes. Typically used to avoid calls to m_pullup() which would otherwise be unnecessary or undesirable. arg is a convenience argument which is passed to the callback function f. Each time f() is called, it will be passed arg, a pointer to the data in the current mbuf, and the length len of the data in this mbuf to which the function should be applied. The function should return zero to indicate success; otherwise, if an error is indi‐ cated, then m_apply() will return the error and stop iterating through the mbuf chain. m_getptr(mbuf, loc, off) Return a pointer to the mbuf containing the data located at loc bytes from the begin‐ ning of the mbuf chain. The corresponding offset into the mbuf will be stored in *off. m_defrag(m0, how) Defragment an mbuf chain, returning the shortest possible chain of mbufs and clusters. If allocation fails and this can not be completed, NULL will be returned and the orig‐ inal chain will be unchanged. Upon success, the original chain will be freed and the new chain will be returned. how should be either M_WAITOK or M_NOWAIT, depending on the caller's preference. This function is especially useful in network drivers, where certain long mbuf chains must be shortened before being added to TX descriptor lists. m_unshare(m0, how) Create a version of the specified mbuf chain whose contents can be safely modified without affecting other users. If allocation fails and this operation can not be com‐ pleted, NULL will be returned. The original mbuf chain is always reclaimed and the reference count of any shared mbuf clusters is decremented. how should be either M_WAITOK or M_NOWAIT, depending on the caller's preference. As a side-effect of this process the returned mbuf chain may be compacted. This function is especially useful in the transmit path of network code, when data must be encrypted or otherwise altered prior to transmission. HARDWARE-ASSISTED CHECKSUM CALCULATION This section currently applies to TCP/IP only. In order to save the host CPU resources, computing checksums is offloaded to the network interface hardware if possible. The m_pkthdr member of the leading mbuf of a packet contains two fields used for that purpose, int csum_flags and int csum_data. The meaning of those fields depends on the direction a packet flows in, and on whether the packet is fragmented. Henceforth, csum_flags or csum_data of a packet will denote the corresponding field of the m_pkthdr member of the leading mbuf in the mbuf chain containing the packet. On output, checksum offloading is attempted after the outgoing interface has been determined for a packet. The interface-specific field ifnet.if_data.ifi_hwassist (see ifnet(9)) is consulted for the capabilities of the interface to assist in computing checksums. The csum_flags field of the packet header is set to indicate which actions the interface is sup‐ posed to perform on it. The actions unsupported by the network interface are done in the software prior to passing the packet down to the interface driver; such actions will never be requested through csum_flags. The flags demanding a particular action from an interface are as follows: CSUM_IP The IP header checksum is to be computed and stored in the corresponding field of the packet. The hardware is expected to know the format of an IP header to determine the offset of the IP checksum field. CSUM_TCP The TCP checksum is to be computed. (See below.) CSUM_UDP The UDP checksum is to be computed. (See below.) Should a TCP or UDP checksum be offloaded to the hardware, the field csum_data will contain the byte offset of the checksum field relative to the end of the IP header. In this case, the checksum field will be initially set by the TCP/IP module to the checksum of the pseudo header defined by the TCP and UDP specifications. On input, an interface indicates the actions it has performed on a packet by setting one or more of the following flags in csum_flags associated with the packet: CSUM_IP_CHECKED The IP header checksum has been computed. CSUM_IP_VALID The IP header has a valid checksum. This flag can appear only in combination with CSUM_IP_CHECKED. CSUM_DATA_VALID The checksum of the data portion of the IP packet has been computed and stored in the field csum_data in network byte order. CSUM_PSEUDO_HDR Can be set only along with CSUM_DATA_VALID to indicate that the IP data checksum found in csum_data allows for the pseudo header defined by the TCP and UDP specifications. Otherwise the checksum of the pseudo header must be calculated by the host CPU and added to csum_data to obtain the final checksum to be used for TCP or UDP val‐ idation purposes. If a particular network interface just indicates success or failure of TCP or UDP checksum validation without returning the exact value of the checksum to the host CPU, its driver can mark CSUM_DATA_VALID and CSUM_PSEUDO_HDR in csum_flags, and set csum_data to 0xFFFF hexadec‐ imal to indicate a valid checksum. It is a peculiarity of the algorithm used that the Internet checksum calculated over any valid packet will be 0xFFFF as long as the original checksum field is included.
STRESS TESTINGWhen running a kernel compiled with the option MBUF_STRESS_TEST, the following sysctl(8)-controlled options may be used to create various failure/extreme cases for testing of network drivers and other parts of the kernel that rely on mbufs. net.inet.ip.mbuf_frag_size Causes ip_output() to fragment outgoing mbuf chains into fragments of the specified size. Setting this variable to 1 is an excellent way to test the long mbuf chain handling ability of network drivers. kern.ipc.m_defragrandomfailures Causes the function m_defrag() to randomly fail, returning NULL. Any piece of code which uses m_defrag() should be tested with this feature.
RETURN VALUESSee above.
SEE ALSOifnet(9), mbuf_tags(9)
HISTORYMbufs appeared in an early version of BSD. Besides being used for network packets, they were used to store various dynamic structures, such as routing table entries, interface addresses, protocol control blocks, etc. In more recent FreeBSD use of mbufs is almost entirely limited to packet storage, with uma(9) zones being used directly to store other network-related memory. Historically, the mbuf allocator has been a special-purpose memory allocator able to run in interrupt contexts and allocating from a special kernel address space map. As of FreeBSD 5.3, the mbuf allocator is a wrapper around uma(9), allowing caching of mbufs, clus‐ ters, and mbuf + cluster pairs in per-CPU caches, as well as bringing other benefits of slab allocation.
BSD November 13, 2012 BSD
AUTHORSThe original mbuf manual page was written by Yar Tikhiy. The uma(9) mbuf allocator was written by Bosko Milekic.
|This manual||Reference||Other manuals|
|refer to||ifnet(9freebsd) | mbuf(9freebsd) | mbuf_tags(9freebsd) | sendfile(2) | sysctl(8) | uma(9freebsd)|