1ab1e2860SLuigi Rizzo /* 217885a7bSLuigi Rizzo * Copyright (C) 2012-2014 Matteo Landi, Luigi Rizzo, Giuseppe Lettieri. All rights reserved. 3ab1e2860SLuigi Rizzo * 4ab1e2860SLuigi Rizzo * Redistribution and use in source and binary forms, with or without 5ab1e2860SLuigi Rizzo * modification, are permitted provided that the following conditions 6ab1e2860SLuigi Rizzo * are met: 7ab1e2860SLuigi Rizzo * 1. Redistributions of source code must retain the above copyright 8ab1e2860SLuigi Rizzo * notice, this list of conditions and the following disclaimer. 9ab1e2860SLuigi Rizzo * 2. Redistributions in binary form must reproduce the above copyright 10ab1e2860SLuigi Rizzo * notice, this list of conditions and the following disclaimer in the 11ab1e2860SLuigi Rizzo * documentation and/or other materials provided with the distribution. 12ab1e2860SLuigi Rizzo * 13ab1e2860SLuigi Rizzo * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 14ab1e2860SLuigi Rizzo * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15ab1e2860SLuigi Rizzo * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16ab1e2860SLuigi Rizzo * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 17ab1e2860SLuigi Rizzo * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18ab1e2860SLuigi Rizzo * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19ab1e2860SLuigi Rizzo * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20ab1e2860SLuigi Rizzo * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21ab1e2860SLuigi Rizzo * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22ab1e2860SLuigi Rizzo * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23ab1e2860SLuigi Rizzo * SUCH DAMAGE. 24ab1e2860SLuigi Rizzo */ 25ab1e2860SLuigi Rizzo 26ab1e2860SLuigi Rizzo /* 27ab1e2860SLuigi Rizzo * $FreeBSD$ 28ab1e2860SLuigi Rizzo * 29ab1e2860SLuigi Rizzo * (New) memory allocator for netmap 30ab1e2860SLuigi Rizzo */ 31ab1e2860SLuigi Rizzo 32ab1e2860SLuigi Rizzo /* 33ab1e2860SLuigi Rizzo * This allocator creates three memory pools: 34ab1e2860SLuigi Rizzo * nm_if_pool for the struct netmap_if 35ab1e2860SLuigi Rizzo * nm_ring_pool for the struct netmap_ring 36ab1e2860SLuigi Rizzo * nm_buf_pool for the packet buffers. 37ab1e2860SLuigi Rizzo * 38ab1e2860SLuigi Rizzo * that contain netmap objects. Each pool is made of a number of clusters, 39ab1e2860SLuigi Rizzo * multiple of a page size, each containing an integer number of objects. 40ab1e2860SLuigi Rizzo * The clusters are contiguous in user space but not in the kernel. 41ab1e2860SLuigi Rizzo * Only nm_buf_pool needs to be dma-able, 42ab1e2860SLuigi Rizzo * but for convenience use the same type of allocator for all. 43ab1e2860SLuigi Rizzo * 44ab1e2860SLuigi Rizzo * Once mapped, the three pools are exported to userspace 45ab1e2860SLuigi Rizzo * as a contiguous block, starting from nm_if_pool. Each 46ab1e2860SLuigi Rizzo * cluster (and pool) is an integral number of pages. 47ab1e2860SLuigi Rizzo * [ . . . ][ . . . . . .][ . . . . . . . . . .] 48ab1e2860SLuigi Rizzo * nm_if nm_ring nm_buf 49ab1e2860SLuigi Rizzo * 50ab1e2860SLuigi Rizzo * The userspace areas contain offsets of the objects in userspace. 51ab1e2860SLuigi Rizzo * When (at init time) we write these offsets, we find out the index 52ab1e2860SLuigi Rizzo * of the object, and from there locate the offset from the beginning 53ab1e2860SLuigi Rizzo * of the region. 54ab1e2860SLuigi Rizzo * 55ab1e2860SLuigi Rizzo * The invididual allocators manage a pool of memory for objects of 56ab1e2860SLuigi Rizzo * the same size. 57ab1e2860SLuigi Rizzo * The pool is split into smaller clusters, whose size is a 58ab1e2860SLuigi Rizzo * multiple of the page size. The cluster size is chosen 59ab1e2860SLuigi Rizzo * to minimize the waste for a given max cluster size 60ab1e2860SLuigi Rizzo * (we do it by brute force, as we have relatively few objects 61ab1e2860SLuigi Rizzo * per cluster). 62ab1e2860SLuigi Rizzo * 63ab1e2860SLuigi Rizzo * Objects are aligned to the cache line (64 bytes) rounding up object 64ab1e2860SLuigi Rizzo * sizes when needed. A bitmap contains the state of each object. 65ab1e2860SLuigi Rizzo * Allocation scans the bitmap; this is done only on attach, so we are not 66ab1e2860SLuigi Rizzo * too worried about performance 67ab1e2860SLuigi Rizzo * 68ab1e2860SLuigi Rizzo * For each allocator we can define (thorugh sysctl) the size and 69ab1e2860SLuigi Rizzo * number of each object. Memory is allocated at the first use of a 70ab1e2860SLuigi Rizzo * netmap file descriptor, and can be freed when all such descriptors 71ab1e2860SLuigi Rizzo * have been released (including unmapping the memory). 72ab1e2860SLuigi Rizzo * If memory is scarce, the system tries to get as much as possible 73ab1e2860SLuigi Rizzo * and the sysctl values reflect the actual allocation. 74ab1e2860SLuigi Rizzo * Together with desired values, the sysctl export also absolute 75ab1e2860SLuigi Rizzo * min and maximum values that cannot be overridden. 76ab1e2860SLuigi Rizzo * 77ab1e2860SLuigi Rizzo * struct netmap_if: 78ab1e2860SLuigi Rizzo * variable size, max 16 bytes per ring pair plus some fixed amount. 79ab1e2860SLuigi Rizzo * 1024 bytes should be large enough in practice. 80ab1e2860SLuigi Rizzo * 81ab1e2860SLuigi Rizzo * In the worst case we have one netmap_if per ring in the system. 82ab1e2860SLuigi Rizzo * 83ab1e2860SLuigi Rizzo * struct netmap_ring 84ab1e2860SLuigi Rizzo * variable size, 8 byte per slot plus some fixed amount. 85ab1e2860SLuigi Rizzo * Rings can be large (e.g. 4k slots, or >32Kbytes). 86ab1e2860SLuigi Rizzo * We default to 36 KB (9 pages), and a few hundred rings. 87ab1e2860SLuigi Rizzo * 88ab1e2860SLuigi Rizzo * struct netmap_buffer 89ab1e2860SLuigi Rizzo * The more the better, both because fast interfaces tend to have 90ab1e2860SLuigi Rizzo * many slots, and because we may want to use buffers to store 91ab1e2860SLuigi Rizzo * packets in userspace avoiding copies. 92ab1e2860SLuigi Rizzo * Must contain a full frame (eg 1518, or more for vlans, jumbo 93ab1e2860SLuigi Rizzo * frames etc.) plus be nicely aligned, plus some NICs restrict 94ab1e2860SLuigi Rizzo * the size to multiple of 1K or so. Default to 2K 95ab1e2860SLuigi Rizzo */ 96ab1e2860SLuigi Rizzo #ifndef _NET_NETMAP_MEM2_H_ 97ab1e2860SLuigi Rizzo #define _NET_NETMAP_MEM2_H_ 98ab1e2860SLuigi Rizzo 99ab1e2860SLuigi Rizzo 100ab1e2860SLuigi Rizzo #define NETMAP_BUF_MAX_NUM 20*4096*2 /* large machine */ 101ab1e2860SLuigi Rizzo 102ab1e2860SLuigi Rizzo #define NETMAP_POOL_MAX_NAMSZ 32 103ab1e2860SLuigi Rizzo 104ab1e2860SLuigi Rizzo 105ab1e2860SLuigi Rizzo enum { 106ab1e2860SLuigi Rizzo NETMAP_IF_POOL = 0, 107ab1e2860SLuigi Rizzo NETMAP_RING_POOL, 108ab1e2860SLuigi Rizzo NETMAP_BUF_POOL, 109ab1e2860SLuigi Rizzo NETMAP_POOLS_NR 110ab1e2860SLuigi Rizzo }; 111ab1e2860SLuigi Rizzo 112ab1e2860SLuigi Rizzo 113ab1e2860SLuigi Rizzo struct netmap_obj_params { 114ab1e2860SLuigi Rizzo u_int size; 115ab1e2860SLuigi Rizzo u_int num; 116ab1e2860SLuigi Rizzo }; 117ab1e2860SLuigi Rizzo struct netmap_obj_pool { 118ab1e2860SLuigi Rizzo char name[NETMAP_POOL_MAX_NAMSZ]; /* name of the allocator */ 119ab1e2860SLuigi Rizzo 120ab1e2860SLuigi Rizzo /* ---------------------------------------------------*/ 121ab1e2860SLuigi Rizzo /* these are only meaningful if the pool is finalized */ 122ab1e2860SLuigi Rizzo /* (see 'finalized' field in netmap_mem_d) */ 123ab1e2860SLuigi Rizzo u_int objtotal; /* actual total number of objects. */ 124ab1e2860SLuigi Rizzo u_int memtotal; /* actual total memory space */ 125ab1e2860SLuigi Rizzo u_int numclusters; /* actual number of clusters */ 126ab1e2860SLuigi Rizzo 127ab1e2860SLuigi Rizzo u_int objfree; /* number of free objects. */ 128ab1e2860SLuigi Rizzo 129ab1e2860SLuigi Rizzo struct lut_entry *lut; /* virt,phys addresses, objtotal entries */ 130ab1e2860SLuigi Rizzo uint32_t *bitmap; /* one bit per buffer, 1 means free */ 131ab1e2860SLuigi Rizzo uint32_t bitmap_slots; /* number of uint32 entries in bitmap */ 132ab1e2860SLuigi Rizzo /* ---------------------------------------------------*/ 133ab1e2860SLuigi Rizzo 134ab1e2860SLuigi Rizzo /* limits */ 135ab1e2860SLuigi Rizzo u_int objminsize; /* minimum object size */ 136ab1e2860SLuigi Rizzo u_int objmaxsize; /* maximum object size */ 137ab1e2860SLuigi Rizzo u_int nummin; /* minimum number of objects */ 138ab1e2860SLuigi Rizzo u_int nummax; /* maximum number of objects */ 139ab1e2860SLuigi Rizzo 140ab1e2860SLuigi Rizzo /* these are changed only by config */ 141ab1e2860SLuigi Rizzo u_int _objtotal; /* total number of objects */ 142ab1e2860SLuigi Rizzo u_int _objsize; /* object size */ 143ab1e2860SLuigi Rizzo u_int _clustsize; /* cluster size */ 144ab1e2860SLuigi Rizzo u_int _clustentries; /* objects per cluster */ 145ab1e2860SLuigi Rizzo u_int _numclusters; /* number of clusters */ 146ab1e2860SLuigi Rizzo 147ab1e2860SLuigi Rizzo /* requested values */ 148ab1e2860SLuigi Rizzo u_int r_objtotal; 149ab1e2860SLuigi Rizzo u_int r_objsize; 150ab1e2860SLuigi Rizzo }; 151ab1e2860SLuigi Rizzo 152ab1e2860SLuigi Rizzo #ifdef linux 153ab1e2860SLuigi Rizzo // XXX a mtx would suffice here 20130415 lr 154ab1e2860SLuigi Rizzo #define NMA_LOCK_T struct semaphore 155ab1e2860SLuigi Rizzo #else /* !linux */ 156ab1e2860SLuigi Rizzo #define NMA_LOCK_T struct mtx 157ab1e2860SLuigi Rizzo #endif /* linux */ 158ab1e2860SLuigi Rizzo 159ab1e2860SLuigi Rizzo typedef int (*netmap_mem_config_t)(struct netmap_mem_d*); 160ab1e2860SLuigi Rizzo typedef int (*netmap_mem_finalize_t)(struct netmap_mem_d*); 161ab1e2860SLuigi Rizzo typedef void (*netmap_mem_deref_t)(struct netmap_mem_d*); 162ab1e2860SLuigi Rizzo 163*f0ea3689SLuigi Rizzo typedef uint16_t nm_memid_t; 164ab1e2860SLuigi Rizzo 165ab1e2860SLuigi Rizzo /* We implement two kinds of netmap_mem_d structures: 166ab1e2860SLuigi Rizzo * 167ab1e2860SLuigi Rizzo * - global: used by hardware NICS; 168ab1e2860SLuigi Rizzo * 169ab1e2860SLuigi Rizzo * - private: used by VALE ports. 170ab1e2860SLuigi Rizzo * 171ab1e2860SLuigi Rizzo * In both cases, the netmap_mem_d structure has the same lifetime as the 172ab1e2860SLuigi Rizzo * netmap_adapter of the corresponding NIC or port. It is the responsibility of 173ab1e2860SLuigi Rizzo * the client code to delete the private allocator when the associated 174ab1e2860SLuigi Rizzo * netmap_adapter is freed (this is implemented by the NAF_MEM_OWNER flag in 175ab1e2860SLuigi Rizzo * netmap.c). The 'refcount' field counts the number of active users of the 176ab1e2860SLuigi Rizzo * structure. The global allocator uses this information to prevent/allow 177ab1e2860SLuigi Rizzo * reconfiguration. The private allocators release all their memory when there 178ab1e2860SLuigi Rizzo * are no active users. By 'active user' we mean an existing netmap_priv 179ab1e2860SLuigi Rizzo * structure holding a reference to the allocator. 180ab1e2860SLuigi Rizzo */ 181ab1e2860SLuigi Rizzo struct netmap_mem_d { 182ab1e2860SLuigi Rizzo NMA_LOCK_T nm_mtx; /* protect the allocator */ 183ab1e2860SLuigi Rizzo u_int nm_totalsize; /* shorthand */ 184ab1e2860SLuigi Rizzo 185ab1e2860SLuigi Rizzo u_int flags; 186ab1e2860SLuigi Rizzo #define NETMAP_MEM_FINALIZED 0x1 /* preallocation done */ 187ab1e2860SLuigi Rizzo #define NETMAP_MEM_PRIVATE 0x2 /* uses private address space */ 188ab1e2860SLuigi Rizzo int lasterr; /* last error for curr config */ 189ab1e2860SLuigi Rizzo int refcount; /* existing priv structures */ 190ab1e2860SLuigi Rizzo /* the three allocators */ 191ab1e2860SLuigi Rizzo struct netmap_obj_pool pools[NETMAP_POOLS_NR]; 192ab1e2860SLuigi Rizzo 193ab1e2860SLuigi Rizzo netmap_mem_config_t config; 194ab1e2860SLuigi Rizzo netmap_mem_finalize_t finalize; 195ab1e2860SLuigi Rizzo netmap_mem_deref_t deref; 196*f0ea3689SLuigi Rizzo 197*f0ea3689SLuigi Rizzo nm_memid_t nm_id; /* allocator identifier */ 198*f0ea3689SLuigi Rizzo 199*f0ea3689SLuigi Rizzo /* list of all existing allocators, sorted by nm_id */ 200*f0ea3689SLuigi Rizzo struct netmap_mem_d *prev, *next; 201ab1e2860SLuigi Rizzo }; 202ab1e2860SLuigi Rizzo 203ab1e2860SLuigi Rizzo extern struct netmap_mem_d nm_mem; 204ab1e2860SLuigi Rizzo 205ab1e2860SLuigi Rizzo vm_paddr_t netmap_mem_ofstophys(struct netmap_mem_d *, vm_ooffset_t); 206ab1e2860SLuigi Rizzo int netmap_mem_finalize(struct netmap_mem_d *); 207ab1e2860SLuigi Rizzo int netmap_mem_init(void); 208ab1e2860SLuigi Rizzo void netmap_mem_fini(void); 209f9790aebSLuigi Rizzo struct netmap_if * 210f9790aebSLuigi Rizzo netmap_mem_if_new(const char *, struct netmap_adapter *); 211f9790aebSLuigi Rizzo void netmap_mem_if_delete(struct netmap_adapter *, struct netmap_if *); 212f9790aebSLuigi Rizzo int netmap_mem_rings_create(struct netmap_adapter *); 213f9790aebSLuigi Rizzo void netmap_mem_rings_delete(struct netmap_adapter *); 214ab1e2860SLuigi Rizzo void netmap_mem_deref(struct netmap_mem_d *); 215*f0ea3689SLuigi Rizzo int netmap_mem_get_info(struct netmap_mem_d *, u_int *size, u_int *memflags, uint16_t *id); 216f9790aebSLuigi Rizzo ssize_t netmap_mem_if_offset(struct netmap_mem_d *, const void *vaddr); 217*f0ea3689SLuigi Rizzo struct netmap_mem_d* netmap_mem_private_new(const char *name, 218*f0ea3689SLuigi Rizzo u_int txr, u_int txd, u_int rxr, u_int rxd, u_int extra_bufs, u_int npipes, 219*f0ea3689SLuigi Rizzo int* error); 220f9790aebSLuigi Rizzo void netmap_mem_private_delete(struct netmap_mem_d *); 221ab1e2860SLuigi Rizzo 222ab1e2860SLuigi Rizzo #define NETMAP_BDG_BUF_SIZE(n) ((n)->pools[NETMAP_BUF_POOL]._objsize) 223ab1e2860SLuigi Rizzo 224*f0ea3689SLuigi Rizzo uint32_t netmap_extra_alloc(struct netmap_adapter *, uint32_t *, uint32_t n); 225ab1e2860SLuigi Rizzo 226ab1e2860SLuigi Rizzo 227ab1e2860SLuigi Rizzo #endif 228