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 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->max_fds <= NR_OPEN_DEFAULT) {
 105                 /*
 106                  * This fdtable is embedded in the files structure and that
 107                  * structure itself is getting destroyed.
 108                  */
 109                 kmem_cache_free(files_cachep,
 110                                 container_of(fdt, struct files_struct, fdtab));
 111                 return;
 112         }
 113         if (fdset_size <= PAGE_SIZE && fdarray_size <= PAGE_SIZE) {
 114                 kfree(fdt->open_fds);
 115                 kfree(fdt->close_on_exec);
 116                 kfree(fdt->fd);
 117                 kfree(fdt);
 118         } else {
 119                 fddef = &get_cpu_var(fdtable_defer_list);
 120                 spin_lock(&fddef->lock);
 121                 fdt->next = fddef->next;
 122                 fddef->next = fdt;
 123                 /* vmallocs are handled from the workqueue context */
 124                 schedule_work(&fddef->wq);
 125                 spin_unlock(&fddef->lock);
 126                 put_cpu_var(fdtable_defer_list);
 127         }
 128 }
 129
 130 /*
 131  * Expand the fdset in the files_struct.  Called with the files spinlock
 132  * held for write.
 133  */
 134 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *fdt)
 135 {
 136         int i;
 137         int count;
 138
 139         BUG_ON(nfdt->max_fds < fdt->max_fds);
 140         /* Copy the existing tables and install the new pointers */
 141
 142         i = fdt->max_fds / (sizeof(unsigned long) * 8);
 143         count = (nfdt->max_fds - fdt->max_fds) / 8;
 144
 145         /*
 146          * Don't copy the entire array if the current fdset is
 147          * not yet initialised.
 148          */
 149         if (i) {
 150                 memcpy (nfdt->open_fds, fdt->open_fds,
 151                                                 fdt->max_fds/8);
 152                 memcpy (nfdt->close_on_exec, fdt->close_on_exec,
 153                                                 fdt->max_fds/8);
 154                 memset (&nfdt->open_fds->fds_bits[i], 0, count);
 155                 memset (&nfdt->close_on_exec->fds_bits[i], 0, count);
 156         }
 157
 158         /* Don't copy/clear the array if we are creating a new
 159            fd array for fork() */
 160         if (fdt->max_fds) {
 161                 memcpy(nfdt->fd, fdt->fd,
 162                         fdt->max_fds * sizeof(struct file *));
 163                 /* clear the remainder of the array */
 164                 memset(&nfdt->fd[fdt->max_fds], 0,
 165                        (nfdt->max_fds - fdt->max_fds) *
 166                                         sizeof(struct file *));
 167         }
 168 }
 169
 170 /*
 171  * Allocate an fdset array, using kmalloc or vmalloc.
 172  * Note: the array isn't cleared at allocation time.
 173  */
 174 fd_set * alloc_fdset(int num)
 175 {
 176         fd_set *new_fdset;
 177         int size = num / 8;
 178
 179         if (size <= PAGE_SIZE)
 180                 new_fdset = (fd_set *) kmalloc(size, GFP_KERNEL);
 181         else
 182                 new_fdset = (fd_set *) vmalloc(size);
 183         return new_fdset;
 184 }
 185
 186 void free_fdset(fd_set *array, int num)
 187 {
 188         if (num <= NR_OPEN_DEFAULT) /* Don't free an embedded fdset */
 189                 return;
 190         else if (num <= 8 * PAGE_SIZE)
 191                 kfree(array);
 192         else
 193                 vfree(array);
 194 }
 195
 196 static struct fdtable *alloc_fdtable(int nr)
 197 {
 198         struct fdtable *fdt = NULL;
 199         int nfds = 0;
 200         fd_set *new_openset = NULL, *new_execset = NULL;
 201         struct file **new_fds;
 202
 203         fdt = kzalloc(sizeof(*fdt), GFP_KERNEL);
 204         if (!fdt)
 205                 goto out;
 206
 207         nfds = NR_OPEN_DEFAULT;
 208         /*
 209          * Expand to the max in easy steps, and keep expanding it until
 210          * we have enough for the requested fd array size.
 211          */
 212         do {
 213 #if NR_OPEN_DEFAULT < 256
 214                 if (nfds < 256)
 215                         nfds = 256;
 216                 else
 217 #endif
 218                 if (nfds < (PAGE_SIZE / sizeof(struct file *)))
 219                         nfds = PAGE_SIZE / sizeof(struct file *);
 220                 else {
 221                         nfds = nfds * 2;
 222                         if (nfds > NR_OPEN)
 223                                 nfds = NR_OPEN;
 224                 }
 225         } while (nfds <= nr);
 226
 227         new_openset = alloc_fdset(nfds);
 228         new_execset = alloc_fdset(nfds);
 229         if (!new_openset || !new_execset)
 230                 goto out;
 231         fdt->open_fds = new_openset;
 232         fdt->close_on_exec = new_execset;
 233
 234         new_fds = alloc_fd_array(nfds);
 235         if (!new_fds)
 236                 goto out;
 237         fdt->fd = new_fds;
 238         fdt->max_fds = nfds;
 239         return fdt;
 240 out:
 241         free_fdset(new_openset, nfds);
 242         free_fdset(new_execset, nfds);
 243         kfree(fdt);
 244         return NULL;
 245 }
 246
 247 /*
 248  * Expand the file descriptor table.
 249  * This function will allocate a new fdtable and both fd array and fdset, of
 250  * the given size.
 251  * Return <0 error code on error; 1 on successful completion.
 252  * The files->file_lock should be held on entry, and will be held on exit.
 253  */
 254 static int expand_fdtable(struct files_struct *files, int nr)
 255         __releases(files->file_lock)
 256         __acquires(files->file_lock)
 257 {
 258         struct fdtable *new_fdt, *cur_fdt;
 259
 260         spin_unlock(&files->file_lock);
 261         new_fdt = alloc_fdtable(nr);
 262         spin_lock(&files->file_lock);
 263         if (!new_fdt)
 264                 return -ENOMEM;
 265         /*
 266          * Check again since another task may have expanded the fd table while
 267          * we dropped the lock
 268          */
 269         cur_fdt = files_fdtable(files);
 270         if (nr >= cur_fdt->max_fds) {
 271                 /* Continue as planned */
 272                 copy_fdtable(new_fdt, cur_fdt);
 273                 rcu_assign_pointer(files->fdt, new_fdt);
 274                 if (cur_fdt->max_fds > NR_OPEN_DEFAULT)
 275                         call_rcu(&cur_fdt->rcu, free_fdtable_rcu);
 276         } else {
 277                 /* Somebody else expanded, so undo our attempt */
 278                 __free_fdtable(new_fdt);
 279         }
 280         return 1;
 281 }
 282
 283 /*
 284  * Expand files.
 285  * This function will expand the file structures, if the requested size exceeds
 286  * the current capacity and there is room for expansion.
 287  * Return <0 error code on error; 0 when nothing done; 1 when files were
 288  * expanded and execution may have blocked.
 289  * The files->file_lock should be held on entry, and will be held on exit.
 290  */
 291 int expand_files(struct files_struct *files, int nr)
 292 {
 293         struct fdtable *fdt;
 294
 295         fdt = files_fdtable(files);
 296         /* Do we need to expand? */
 297         if (nr < fdt->max_fds)
 298                 return 0;
 299         /* Can we expand? */
 300         if (nr >= NR_OPEN)
 301                 return -EMFILE;
 302
 303         /* All good, so we try */
 304         return expand_fdtable(files, nr);
 305 }
 306
 307 static void __devinit fdtable_defer_list_init(int cpu)
 308 {
 309         struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
 310         spin_lock_init(&fddef->lock);
 311         INIT_WORK(&fddef->wq, free_fdtable_work);
 312         fddef->next = NULL;
 313 }
 314
 315 void __init files_defer_init(void)
 316 {
 317         int i;
 318         for_each_possible_cpu(i)
 319                 fdtable_defer_list_init(i);
 320 }