fs/file.c

   1 /*
   2  *  linux/fs/file.c
   3  *
   4  *  Copyright (C) 1998-1999, Stephen Tweedie and Bill Hawes
   5  *
   6  *  Manage the dynamic fd arrays in the process files_struct.
   7  */
   8
   9 #include <linux/fs.h>
  10 #include <linux/mm.h>
  11 #include <linux/time.h>
  12 #include <linux/slab.h>
  13 #include <linux/vmalloc.h>
  14 #include <linux/file.h>
  15 #include <linux/bitops.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/spinlock.h>
  18 #include <linux/rcupdate.h>
  19 #include <linux/workqueue.h>
  20
  21 struct fdtable_defer {
  22         spinlock_t lock;
  23         struct work_struct wq;
  24         struct fdtable *next;
  25 };
  26
  27 /*
  28  * We use this list to defer free fdtables that have vmalloced
  29  * sets/arrays. By keeping a per-cpu list, we avoid having to embed
  30  * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
  31  * this per-task structure.
  32  */
  33 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
  34
  35
  36 /*
  37  * Allocate an fd array, using kmalloc or vmalloc.
  38  * Note: the array isn't cleared at allocation time.
  39  */
  40 struct file ** alloc_fd_array(int num)
  41 {
  42         struct file **new_fds;
  43         int size = num * sizeof(struct file *);
  44
  45         if (size <= PAGE_SIZE)
  46                 new_fds = (struct file **) kmalloc(size, GFP_KERNEL);
  47         else
  48                 new_fds = (struct file **) vmalloc(size);
  49         return new_fds;
  50 }
  51
  52 void free_fd_array(struct file **array, int num)
  53 {
  54         int size = num * sizeof(struct file *);
  55
  56         if (!array) {
  57                 printk (KERN_ERR "free_fd_array: array = 0 (num = %d)\n", num);
  58                 return;
  59         }
  60
  61         if (num <= NR_OPEN_DEFAULT) /* Don't free the embedded fd array! */
  62                 return;
  63         else if (size <= PAGE_SIZE)
  64                 kfree(array);
  65         else
  66                 vfree(array);
  67 }
  68
  69 static void __free_fdtable(struct fdtable *fdt)
  70 {
  71         free_fdset(fdt->open_fds, fdt->max_fds);
  72         free_fdset(fdt->close_on_exec, fdt->max_fds);
  73         free_fd_array(fdt->fd, fdt->max_fds);
  74         kfree(fdt);
  75 }
  76
  77 static void free_fdtable_work(struct work_struct *work)
  78 {
  79         struct fdtable_defer *f =
  80                 container_of(work, struct fdtable_defer, wq);
  81         struct fdtable *fdt;
  82
  83         spin_lock_bh(&f->lock);
  84         fdt = f->next;
  85         f->next = NULL;
  86         spin_unlock_bh(&f->lock);
  87         while(fdt) {
  88                 struct fdtable *next = fdt->next;
  89                 __free_fdtable(fdt);
  90                 fdt = next;
  91         }
  92 }
  93
  94 static void free_fdtable_rcu(struct rcu_head *rcu)
  95 {
  96         struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
  97         int fdset_size, fdarray_size;
  98         struct fdtable_defer *fddef;
  99
 100         BUG_ON(!fdt);
 101         fdset_size = fdt->max_fds / 8;
 102         fdarray_size = fdt->max_fds * sizeof(struct file *);
 103
 104         if (fdt->free_files) {
 105                 /*
 106                  * The this fdtable was embedded in the files structure
 107                  * and the files structure itself was getting destroyed.
 108                  * It is now safe to free the files structure.
 109                  */
 110                 kmem_cache_free(files_cachep, fdt->free_files);
 111                 return;
 112         }
 113         if (fdt->max_fds <= NR_OPEN_DEFAULT)
 114                 /*
 115                  * The fdtable was embedded
 116                  */
 117                 return;
 118         if (fdset_size <= PAGE_SIZE && fdarray_size <= PAGE_SIZE) {
 119                 kfree(fdt->open_fds);
 120                 kfree(fdt->close_on_exec);
 121                 kfree(fdt->fd);
 122                 kfree(fdt);
 123         } else {
 124                 fddef = &get_cpu_var(fdtable_defer_list);
 125                 spin_lock(&fddef->lock);
 126                 fdt->next = fddef->next;
 127                 fddef->next = fdt;
 128                 /* vmallocs are handled from the workqueue context */
 129                 schedule_work(&fddef->wq);
 130                 spin_unlock(&fddef->lock);
 131                 put_cpu_var(fdtable_defer_list);
 132         }
 133 }
 134
 135 void free_fdtable(struct fdtable *fdt)
 136 {
 137         if (fdt->free_files || fdt->max_fds > NR_OPEN_DEFAULT)
 138                 call_rcu(&fdt->rcu, free_fdtable_rcu);
 139 }
 140
 141 /*
 142  * Expand the fdset in the files_struct.  Called with the files spinlock
 143  * held for write.
 144  */
 145 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *fdt)
 146 {
 147         int i;
 148         int count;
 149
 150         BUG_ON(nfdt->max_fds < fdt->max_fds);
 151         /* Copy the existing tables and install the new pointers */
 152
 153         i = fdt->max_fds / (sizeof(unsigned long) * 8);
 154         count = (nfdt->max_fds - fdt->max_fds) / 8;
 155
 156         /*
 157          * Don't copy the entire array if the current fdset is
 158          * not yet initialised.
 159          */
 160         if (i) {
 161                 memcpy (nfdt->open_fds, fdt->open_fds,
 162                                                 fdt->max_fds/8);
 163                 memcpy (nfdt->close_on_exec, fdt->close_on_exec,
 164                                                 fdt->max_fds/8);
 165                 memset (&nfdt->open_fds->fds_bits[i], 0, count);
 166                 memset (&nfdt->close_on_exec->fds_bits[i], 0, count);
 167         }
 168
 169         /* Don't copy/clear the array if we are creating a new
 170            fd array for fork() */
 171         if (fdt->max_fds) {
 172                 memcpy(nfdt->fd, fdt->fd,
 173                         fdt->max_fds * sizeof(struct file *));
 174                 /* clear the remainder of the array */
 175                 memset(&nfdt->fd[fdt->max_fds], 0,
 176                        (nfdt->max_fds - fdt->max_fds) *
 177                                         sizeof(struct file *));
 178         }
 179 }
 180
 181 /*
 182  * Allocate an fdset array, using kmalloc or vmalloc.
 183  * Note: the array isn't cleared at allocation time.
 184  */
 185 fd_set * alloc_fdset(int num)
 186 {
 187         fd_set *new_fdset;
 188         int size = num / 8;
 189
 190         if (size <= PAGE_SIZE)
 191                 new_fdset = (fd_set *) kmalloc(size, GFP_KERNEL);
 192         else
 193                 new_fdset = (fd_set *) vmalloc(size);
 194         return new_fdset;
 195 }
 196
 197 void free_fdset(fd_set *array, int num)
 198 {
 199         if (num <= NR_OPEN_DEFAULT) /* Don't free an embedded fdset */
 200                 return;
 201         else if (num <= 8 * PAGE_SIZE)
 202                 kfree(array);
 203         else
 204                 vfree(array);
 205 }
 206
 207 static struct fdtable *alloc_fdtable(int nr)
 208 {
 209         struct fdtable *fdt = NULL;
 210         int nfds = 0;
 211         fd_set *new_openset = NULL, *new_execset = NULL;
 212         struct file **new_fds;
 213
 214         fdt = kzalloc(sizeof(*fdt), GFP_KERNEL);
 215         if (!fdt)
 216                 goto out;
 217
 218         nfds = NR_OPEN_DEFAULT;
 219         /*
 220          * Expand to the max in easy steps, and keep expanding it until
 221          * we have enough for the requested fd array size.
 222          */
 223         do {
 224 #if NR_OPEN_DEFAULT < 256
 225                 if (nfds < 256)
 226                         nfds = 256;
 227                 else
 228 #endif
 229                 if (nfds < (PAGE_SIZE / sizeof(struct file *)))
 230                         nfds = PAGE_SIZE / sizeof(struct file *);
 231                 else {
 232                         nfds = nfds * 2;
 233                         if (nfds > NR_OPEN)
 234                                 nfds = NR_OPEN;
 235                 }
 236         } while (nfds <= nr);
 237
 238         new_openset = alloc_fdset(nfds);
 239         new_execset = alloc_fdset(nfds);
 240         if (!new_openset || !new_execset)
 241                 goto out;
 242         fdt->open_fds = new_openset;
 243         fdt->close_on_exec = new_execset;
 244
 245         new_fds = alloc_fd_array(nfds);
 246         if (!new_fds)
 247                 goto out;
 248         fdt->fd = new_fds;
 249         fdt->max_fds = nfds;
 250         fdt->free_files = NULL;
 251         return fdt;
 252 out:
 253         free_fdset(new_openset, nfds);
 254         free_fdset(new_execset, nfds);
 255         kfree(fdt);
 256         return NULL;
 257 }
 258
 259 /*
 260  * Expand the file descriptor table.
 261  * This function will allocate a new fdtable and both fd array and fdset, of
 262  * the given size.
 263  * Return <0 error code on error; 1 on successful completion.
 264  * The files->file_lock should be held on entry, and will be held on exit.
 265  */
 266 static int expand_fdtable(struct files_struct *files, int nr)
 267         __releases(files->file_lock)
 268         __acquires(files->file_lock)
 269 {
 270         struct fdtable *new_fdt, *cur_fdt;
 271
 272         spin_unlock(&files->file_lock);
 273         new_fdt = alloc_fdtable(nr);
 274         spin_lock(&files->file_lock);
 275         if (!new_fdt)
 276                 return -ENOMEM;
 277         /*
 278          * Check again since another task may have expanded the fd table while
 279          * we dropped the lock
 280          */
 281         cur_fdt = files_fdtable(files);
 282         if (nr >= cur_fdt->max_fds) {
 283                 /* Continue as planned */
 284                 copy_fdtable(new_fdt, cur_fdt);
 285                 rcu_assign_pointer(files->fdt, new_fdt);
 286                 free_fdtable(cur_fdt);
 287         } else {
 288                 /* Somebody else expanded, so undo our attempt */
 289                 __free_fdtable(new_fdt);
 290         }
 291         return 1;
 292 }
 293
 294 /*
 295  * Expand files.
 296  * This function will expand the file structures, if the requested size exceeds
 297  * the current capacity and there is room for expansion.
 298  * Return <0 error code on error; 0 when nothing done; 1 when files were
 299  * expanded and execution may have blocked.
 300  * The files->file_lock should be held on entry, and will be held on exit.
 301  */
 302 int expand_files(struct files_struct *files, int nr)
 303 {
 304         struct fdtable *fdt;
 305
 306         fdt = files_fdtable(files);
 307         /* Do we need to expand? */
 308         if (nr < fdt->max_fds)
 309                 return 0;
 310         /* Can we expand? */
 311         if (nr >= NR_OPEN)
 312                 return -EMFILE;
 313
 314         /* All good, so we try */
 315         return expand_fdtable(files, nr);
 316 }
 317
 318 static void __devinit fdtable_defer_list_init(int cpu)
 319 {
 320         struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
 321         spin_lock_init(&fddef->lock);
 322         INIT_WORK(&fddef->wq, free_fdtable_work);
 323         fddef->next = NULL;
 324 }
 325
 326 void __init files_defer_init(void)
 327 {
 328         int i;
 329         for_each_possible_cpu(i)
 330                 fdtable_defer_list_init(i);
 331 }