sys/vm/uma_int.h

   1 /*-
   2  * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
   3  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
   4  * All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice unmodified, this list of conditions, and the following
  11  *    disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
  20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  21  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  22  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  23  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  24  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
  25  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  26  *
  27  * $FreeBSD$
  28  *
  29  */
  30
  31 #include <sys/_task.h>
  32
  33 /*
  34  * This file includes definitions, structures, prototypes, and inlines that
  35  * should not be used outside of the actual implementation of UMA.
  36  */
  37
  38 /*
  39  * Here's a quick description of the relationship between the objects:
  40  *
  41  * Kegs contain lists of slabs which are stored in either the full bin, empty
  42  * bin, or partially allocated bin, to reduce fragmentation.  They also contain
  43  * the user supplied value for size, which is adjusted for alignment purposes
  44  * and rsize is the result of that.  The Keg also stores information for
  45  * managing a hash of page addresses that maps pages to uma_slab_t structures
  46  * for pages that don't have embedded uma_slab_t's.
  47  *
  48  * The uma_slab_t may be embedded in a UMA_SLAB_SIZE chunk of memory or it may
  49  * be allocated off the page from a special slab zone.  The free list within a
  50  * slab is managed with a bitmask.  For item sizes that would yield more than
  51  * 10% memory waste we potentially allocate a separate uma_slab_t if this will
  52  * improve the number of items per slab that will fit.
  53  *
  54  * The only really gross cases, with regards to memory waste, are for those
  55  * items that are just over half the page size.   You can get nearly 50% waste,
  56  * so you fall back to the memory footprint of the power of two allocator. I
  57  * have looked at memory allocation sizes on many of the machines available to
  58  * me, and there does not seem to be an abundance of allocations at this range
  59  * so at this time it may not make sense to optimize for it.  This can, of
  60  * course, be solved with dynamic slab sizes.
  61  *
  62  * Kegs may serve multiple Zones but by far most of the time they only serve
  63  * one.  When a Zone is created, a Keg is allocated and setup for it.  While
  64  * the backing Keg stores slabs, the Zone caches Buckets of items allocated
  65  * from the slabs.  Each Zone is equipped with an init/fini and ctor/dtor
  66  * pair, as well as with its own set of small per-CPU caches, layered above
  67  * the Zone's general Bucket cache.
  68  *
  69  * The PCPU caches are protected by critical sections, and may be accessed
  70  * safely only from their associated CPU, while the Zones backed by the same
  71  * Keg all share a common Keg lock (to coalesce contention on the backing
  72  * slabs).  The backing Keg typically only serves one Zone but in the case of
  73  * multiple Zones, one of the Zones is considered the Master Zone and all
  74  * Zone-related stats from the Keg are done in the Master Zone.  For an
  75  * example of a Multi-Zone setup, refer to the Mbuf allocation code.
  76  */
  77
  78 /*
  79  *      This is the representation for normal (Non OFFPAGE slab)
  80  *
  81  *      i == item
  82  *      s == slab pointer
  83  *
  84  *      <----------------  Page (UMA_SLAB_SIZE) ------------------>
  85  *      ___________________________________________________________
  86  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   ___________ |
  87  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header||
  88  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________||
  89  *     |___________________________________________________________|
  90  *
  91  *
  92  *      This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE.
  93  *
  94  *      ___________________________________________________________
  95  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   |
  96  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i|  |
  97  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_|  |
  98  *     |___________________________________________________________|
  99  *       ___________    ^
 100  *      |slab header|   |
 101  *      |___________|---*
 102  *
 103  */
 104
 105 #ifndef VM_UMA_INT_H
 106 #define VM_UMA_INT_H
 107
 108 #define UMA_SLAB_SIZE   PAGE_SIZE       /* How big are our slabs? */
 109 #define UMA_SLAB_MASK   (PAGE_SIZE - 1) /* Mask to get back to the page */
 110 #define UMA_SLAB_SHIFT  PAGE_SHIFT      /* Number of bits PAGE_MASK */
 111
 112 #define UMA_BOOT_PAGES          64      /* Pages allocated for startup */
 113 #define UMA_BOOT_PAGES_ZONES    32      /* Multiplier for pages to reserve */
 114                                         /* if uma_zone > PAGE_SIZE */
 115
 116 /* Max waste percentage before going to off page slab management */
 117 #define UMA_MAX_WASTE   10
 118
 119 /*
 120  * I doubt there will be many cases where this is exceeded. This is the initial
 121  * size of the hash table for uma_slabs that are managed off page. This hash
 122  * does expand by powers of two.  Currently it doesn't get smaller.
 123  */
 124 #define UMA_HASH_SIZE_INIT      32
 125
 126 /*
 127  * I should investigate other hashing algorithms.  This should yield a low
 128  * number of collisions if the pages are relatively contiguous.
 129  */
 130
 131 #define UMA_HASH(h, s) ((((uintptr_t)s) >> UMA_SLAB_SHIFT) & (h)->uh_hashmask)
 132
 133 #define UMA_HASH_INSERT(h, s, mem)                                      \
 134                 SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h),      \
 135                     (mem))], (s), us_hlink)
 136 #define UMA_HASH_REMOVE(h, s, mem)                                      \
 137                 SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h),           \
 138                     (mem))], (s), uma_slab, us_hlink)
 139
 140 /* Hash table for freed address -> slab translation */
 141
 142 SLIST_HEAD(slabhead, uma_slab);
 143
 144 struct uma_hash {
 145         struct slabhead *uh_slab_hash;  /* Hash table for slabs */
 146         int             uh_hashsize;    /* Current size of the hash table */
 147         int             uh_hashmask;    /* Mask used during hashing */
 148 };
 149
 150 /*
 151  * align field or structure to cache line
 152  */
 153 #if defined(__amd64__)
 154 #define UMA_ALIGN       __aligned(CACHE_LINE_SIZE)
 155 #else
 156 #define UMA_ALIGN
 157 #endif
 158
 159 /*
 160  * Structures for per cpu queues.
 161  */
 162
 163 struct uma_bucket {
 164         LIST_ENTRY(uma_bucket)  ub_link;        /* Link into the zone */
 165         int16_t ub_cnt;                         /* Count of free items. */
 166         int16_t ub_entries;                     /* Max items. */
 167         void    *ub_bucket[];                   /* actual allocation storage */
 168 };
 169
 170 typedef struct uma_bucket * uma_bucket_t;
 171
 172 struct uma_cache {
 173         uma_bucket_t    uc_freebucket;  /* Bucket we're freeing to */
 174         uma_bucket_t    uc_allocbucket; /* Bucket to allocate from */
 175         uint64_t        uc_allocs;      /* Count of allocations */
 176         uint64_t        uc_frees;       /* Count of frees */
 177 } UMA_ALIGN;
 178
 179 typedef struct uma_cache * uma_cache_t;
 180
 181 /*
 182  * Keg management structure
 183  *
 184  * TODO: Optimize for cache line size
 185  *
 186  */
 187 struct uma_keg {
 188         struct mtx_padalign     uk_lock;        /* Lock for the keg */
 189         struct uma_hash uk_hash;
 190
 191         LIST_HEAD(,uma_zone)    uk_zones;       /* Keg's zones */
 192         LIST_HEAD(,uma_slab)    uk_part_slab;   /* partially allocated slabs */
 193         LIST_HEAD(,uma_slab)    uk_free_slab;   /* empty slab list */
 194         LIST_HEAD(,uma_slab)    uk_full_slab;   /* full slabs */
 195
 196         uint32_t        uk_align;       /* Alignment mask */
 197         uint32_t        uk_pages;       /* Total page count */
 198         uint32_t        uk_free;        /* Count of items free in slabs */
 199         uint32_t        uk_reserve;     /* Number of reserved items. */
 200         uint32_t        uk_size;        /* Requested size of each item */
 201         uint32_t        uk_rsize;       /* Real size of each item */
 202         uint32_t        uk_maxpages;    /* Maximum number of pages to alloc */
 203
 204         uma_init        uk_init;        /* Keg's init routine */
 205         uma_fini        uk_fini;        /* Keg's fini routine */
 206         uma_alloc       uk_allocf;      /* Allocation function */
 207         uma_free        uk_freef;       /* Free routine */
 208
 209         u_long          uk_offset;      /* Next free offset from base KVA */
 210         vm_offset_t     uk_kva;         /* Zone base KVA */
 211         uma_zone_t      uk_slabzone;    /* Slab zone backing us, if OFFPAGE */
 212
 213         uint16_t        uk_slabsize;    /* Slab size for this keg */
 214         uint16_t        uk_pgoff;       /* Offset to uma_slab struct */
 215         uint16_t        uk_ppera;       /* pages per allocation from backend */
 216         uint16_t        uk_ipers;       /* Items per slab */
 217         uint32_t        uk_flags;       /* Internal flags */
 218
 219         /* Least used fields go to the last cache line. */
 220         const char      *uk_name;               /* Name of creating zone. */
 221         LIST_ENTRY(uma_keg)     uk_link;        /* List of all kegs */
 222 };
 223 typedef struct uma_keg  * uma_keg_t;
 224
 225 /*
 226  * Free bits per-slab.
 227  */
 228 #define SLAB_SETSIZE    (PAGE_SIZE / UMA_SMALLEST_UNIT)
 229 BITSET_DEFINE(slabbits, SLAB_SETSIZE);
 230
 231 /*
 232  * The slab structure manages a single contiguous allocation from backing
 233  * store and subdivides it into individually allocatable items.
 234  */
 235 struct uma_slab {
 236         uma_keg_t       us_keg;                 /* Keg we live in */
 237         union {
 238                 LIST_ENTRY(uma_slab)    _us_link;       /* slabs in zone */
 239                 unsigned long   _us_size;       /* Size of allocation */
 240         } us_type;
 241         SLIST_ENTRY(uma_slab)   us_hlink;       /* Link for hash table */
 242         uint8_t         *us_data;               /* First item */
 243         struct slabbits us_free;                /* Free bitmask. */
 244 #ifdef INVARIANTS
 245         struct slabbits us_debugfree;           /* Debug bitmask. */
 246 #endif
 247         uint16_t        us_freecount;           /* How many are free? */
 248         uint8_t         us_flags;               /* Page flags see uma.h */
 249         uint8_t         us_pad;                 /* Pad to 32bits, unused. */
 250 };
 251
 252 #define us_link us_type._us_link
 253 #define us_size us_type._us_size
 254
 255 typedef struct uma_slab * uma_slab_t;
 256 typedef uma_slab_t (*uma_slaballoc)(uma_zone_t, uma_keg_t, int);
 257
 258 struct uma_klink {
 259         LIST_ENTRY(uma_klink)   kl_link;
 260         uma_keg_t               kl_keg;
 261 };
 262 typedef struct uma_klink *uma_klink_t;
 263
 264 /*
 265  * Zone management structure
 266  *
 267  * TODO: Optimize for cache line size
 268  *
 269  */
 270 struct uma_zone {
 271         struct mtx_padalign     uz_lock;        /* Lock for the zone */
 272         struct mtx_padalign     *uz_lockptr;
 273         const char              *uz_name;       /* Text name of the zone */
 274
 275         LIST_ENTRY(uma_zone)    uz_link;        /* List of all zones in keg */
 276         LIST_HEAD(,uma_bucket)  uz_buckets;     /* full buckets */
 277
 278         LIST_HEAD(,uma_klink)   uz_kegs;        /* List of kegs. */
 279         struct uma_klink        uz_klink;       /* klink for first keg. */
 280
 281         uma_slaballoc   uz_slab;        /* Allocate a slab from the backend. */
 282         uma_ctor        uz_ctor;        /* Constructor for each allocation */
 283         uma_dtor        uz_dtor;        /* Destructor */
 284         uma_init        uz_init;        /* Initializer for each item */
 285         uma_fini        uz_fini;        /* Finalizer for each item. */
 286         uma_import      uz_import;      /* Import new memory to cache. */
 287         uma_release     uz_release;     /* Release memory from cache. */
 288         void            *uz_arg;        /* Import/release argument. */
 289
 290         uint32_t        uz_flags;       /* Flags inherited from kegs */
 291         uint32_t        uz_size;        /* Size inherited from kegs */
 292
 293         volatile u_long uz_allocs UMA_ALIGN; /* Total number of allocations */
 294         volatile u_long uz_fails;       /* Total number of alloc failures */
 295         volatile u_long uz_frees;       /* Total number of frees */
 296         uint64_t        uz_sleeps;      /* Total number of alloc sleeps */
 297         uint16_t        uz_count;       /* Amount of items in full bucket */
 298         uint16_t        uz_count_min;   /* Minimal amount of items there */
 299
 300         /* The next two fields are used to print a rate-limited warnings. */
 301         const char      *uz_warning;    /* Warning to print on failure */
 302         struct timeval  uz_ratecheck;   /* Warnings rate-limiting */
 303
 304         struct task     uz_maxaction;   /* Task to run when at limit */
 305
 306         /*
 307          * This HAS to be the last item because we adjust the zone size
 308          * based on NCPU and then allocate the space for the zones.
 309          */
 310         struct uma_cache        uz_cpu[1]; /* Per cpu caches */
 311 };
 312
 313 /*
 314  * These flags must not overlap with the UMA_ZONE flags specified in uma.h.
 315  */
 316 #define UMA_ZFLAG_MULTI         0x04000000      /* Multiple kegs in the zone. */
 317 #define UMA_ZFLAG_DRAINING      0x08000000      /* Running zone_drain. */
 318 #define UMA_ZFLAG_BUCKET        0x10000000      /* Bucket zone. */
 319 #define UMA_ZFLAG_INTERNAL      0x20000000      /* No offpage no PCPU. */
 320 #define UMA_ZFLAG_FULL          0x40000000      /* Reached uz_maxpages */
 321 #define UMA_ZFLAG_CACHEONLY     0x80000000      /* Don't ask VM for buckets. */
 322
 323 #define UMA_ZFLAG_INHERIT                                               \
 324     (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY | UMA_ZFLAG_BUCKET)
 325
 326 static inline uma_keg_t
 327 zone_first_keg(uma_zone_t zone)
 328 {
 329         uma_klink_t klink;
 330
 331         klink = LIST_FIRST(&zone->uz_kegs);
 332         return (klink != NULL) ? klink->kl_keg : NULL;
 333 }
 334
 335 #undef UMA_ALIGN
 336
 337 #ifdef _KERNEL
 338 /* Internal prototypes */
 339 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, uint8_t *data);
 340 void *uma_large_malloc(vm_size_t size, int wait);
 341 void uma_large_free(uma_slab_t slab);
 342
 343 /* Lock Macros */
 344
 345 #define KEG_LOCK_INIT(k, lc)                                    \
 346         do {                                                    \
 347                 if ((lc))                                       \
 348                         mtx_init(&(k)->uk_lock, (k)->uk_name,   \
 349                             (k)->uk_name, MTX_DEF | MTX_DUPOK); \
 350                 else                                            \
 351                         mtx_init(&(k)->uk_lock, (k)->uk_name,   \
 352                             "UMA zone", MTX_DEF | MTX_DUPOK);   \
 353         } while (0)
 354
 355 #define KEG_LOCK_FINI(k)        mtx_destroy(&(k)->uk_lock)
 356 #define KEG_LOCK(k)     mtx_lock(&(k)->uk_lock)
 357 #define KEG_UNLOCK(k)   mtx_unlock(&(k)->uk_lock)
 358
 359 #define ZONE_LOCK_INIT(z, lc)                                   \
 360         do {                                                    \
 361                 if ((lc))                                       \
 362                         mtx_init(&(z)->uz_lock, (z)->uz_name,   \
 363                             (z)->uz_name, MTX_DEF | MTX_DUPOK); \
 364                 else                                            \
 365                         mtx_init(&(z)->uz_lock, (z)->uz_name,   \
 366                             "UMA zone", MTX_DEF | MTX_DUPOK);   \
 367         } while (0)
 368
 369 #define ZONE_LOCK(z)    mtx_lock((z)->uz_lockptr)
 370 #define ZONE_TRYLOCK(z) mtx_trylock((z)->uz_lockptr)
 371 #define ZONE_UNLOCK(z)  mtx_unlock((z)->uz_lockptr)
 372 #define ZONE_LOCK_FINI(z)       mtx_destroy(&(z)->uz_lock)
 373
 374 /*
 375  * Find a slab within a hash table.  This is used for OFFPAGE zones to lookup
 376  * the slab structure.
 377  *
 378  * Arguments:
 379  *      hash  The hash table to search.
 380  *      data  The base page of the item.
 381  *
 382  * Returns:
 383  *      A pointer to a slab if successful, else NULL.
 384  */
 385 static __inline uma_slab_t
 386 hash_sfind(struct uma_hash *hash, uint8_t *data)
 387 {
 388         uma_slab_t slab;
 389         int hval;
 390
 391         hval = UMA_HASH(hash, data);
 392
 393         SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) {
 394                 if ((uint8_t *)slab->us_data == data)
 395                         return (slab);
 396         }
 397         return (NULL);
 398 }
 399
 400 static __inline uma_slab_t
 401 vtoslab(vm_offset_t va)
 402 {
 403         vm_page_t p;
 404
 405         p = PHYS_TO_VM_PAGE(pmap_kextract(va));
 406         return ((uma_slab_t)p->plinks.s.pv);
 407 }
 408
 409 static __inline void
 410 vsetslab(vm_offset_t va, uma_slab_t slab)
 411 {
 412         vm_page_t p;
 413
 414         p = PHYS_TO_VM_PAGE(pmap_kextract(va));
 415         p->plinks.s.pv = slab;
 416 }
 417
 418 /*
 419  * The following two functions may be defined by architecture specific code
 420  * if they can provide more efficient allocation functions.  This is useful
 421  * for using direct mapped addresses.
 422  */
 423 void *uma_small_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag,
 424     int wait);
 425 void uma_small_free(void *mem, vm_size_t size, uint8_t flags);
 426 #endif /* _KERNEL */
 427
 428 #endif /* VM_UMA_INT_H */