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 int sysctl_nr_open __read_mostly = 1024*1024;
  28
  29 /*
  30  * We use this list to defer free fdtables that have vmalloced
  31  * sets/arrays. By keeping a per-cpu list, we avoid having to embed
  32  * the work_struct in fdtable itself which avoids a 64 byte (i386) increase in
  33  * this per-task structure.
  34  */
  35 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
  36
  37 static inline void * alloc_fdmem(unsigned int size)
  38 {
  39         if (size <= PAGE_SIZE)
  40                 return kmalloc(size, GFP_KERNEL);
  41         else
  42                 return vmalloc(size);
  43 }
  44
  45 static inline void free_fdarr(struct fdtable *fdt)
  46 {
  47         if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *)))
  48                 kfree(fdt->fd);
  49         else
  50                 vfree(fdt->fd);
  51 }
  52
  53 static inline void free_fdset(struct fdtable *fdt)
  54 {
  55         if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2))
  56                 kfree(fdt->open_fds);
  57         else
  58                 vfree(fdt->open_fds);
  59 }
  60
  61 static void free_fdtable_work(struct work_struct *work)
  62 {
  63         struct fdtable_defer *f =
  64                 container_of(work, struct fdtable_defer, wq);
  65         struct fdtable *fdt;
  66
  67         spin_lock_bh(&f->lock);
  68         fdt = f->next;
  69         f->next = NULL;
  70         spin_unlock_bh(&f->lock);
  71         while(fdt) {
  72                 struct fdtable *next = fdt->next;
  73                 vfree(fdt->fd);
  74                 free_fdset(fdt);
  75                 kfree(fdt);
  76                 fdt = next;
  77         }
  78 }
  79
  80 void free_fdtable_rcu(struct rcu_head *rcu)
  81 {
  82         struct fdtable *fdt = container_of(rcu, struct fdtable, rcu);
  83         struct fdtable_defer *fddef;
  84
  85         BUG_ON(!fdt);
  86
  87         if (fdt->max_fds <= NR_OPEN_DEFAULT) {
  88                 /*
  89                  * This fdtable is embedded in the files structure and that
  90                  * structure itself is getting destroyed.
  91                  */
  92                 kmem_cache_free(files_cachep,
  93                                 container_of(fdt, struct files_struct, fdtab));
  94                 return;
  95         }
  96         if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) {
  97                 kfree(fdt->fd);
  98                 kfree(fdt->open_fds);
  99                 kfree(fdt);
 100         } else {
 101                 fddef = &get_cpu_var(fdtable_defer_list);
 102                 spin_lock(&fddef->lock);
 103                 fdt->next = fddef->next;
 104                 fddef->next = fdt;
 105                 /* vmallocs are handled from the workqueue context */
 106                 schedule_work(&fddef->wq);
 107                 spin_unlock(&fddef->lock);
 108                 put_cpu_var(fdtable_defer_list);
 109         }
 110 }
 111
 112 /*
 113  * Expand the fdset in the files_struct.  Called with the files spinlock
 114  * held for write.
 115  */
 116 static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
 117 {
 118         unsigned int cpy, set;
 119
 120         BUG_ON(nfdt->max_fds < ofdt->max_fds);
 121         if (ofdt->max_fds == 0)
 122                 return;
 123
 124         cpy = ofdt->max_fds * sizeof(struct file *);
 125         set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
 126         memcpy(nfdt->fd, ofdt->fd, cpy);
 127         memset((char *)(nfdt->fd) + cpy, 0, set);
 128
 129         cpy = ofdt->max_fds / BITS_PER_BYTE;
 130         set = (nfdt->max_fds - ofdt->max_fds) / BITS_PER_BYTE;
 131         memcpy(nfdt->open_fds, ofdt->open_fds, cpy);
 132         memset((char *)(nfdt->open_fds) + cpy, 0, set);
 133         memcpy(nfdt->close_on_exec, ofdt->close_on_exec, cpy);
 134         memset((char *)(nfdt->close_on_exec) + cpy, 0, set);
 135 }
 136
 137 static struct fdtable * alloc_fdtable(unsigned int nr)
 138 {
 139         struct fdtable *fdt;
 140         char *data;
 141
 142         /*
 143          * Figure out how many fds we actually want to support in this fdtable.
 144          * Allocation steps are keyed to the size of the fdarray, since it
 145          * grows far faster than any of the other dynamic data. We try to fit
 146          * the fdarray into comfortable page-tuned chunks: starting at 1024B
 147          * and growing in powers of two from there on.
 148          */
 149         nr /= (1024 / sizeof(struct file *));
 150         nr = roundup_pow_of_two(nr + 1);
 151         nr *= (1024 / sizeof(struct file *));
 152         if (nr > sysctl_nr_open)
 153                 nr = sysctl_nr_open;
 154
 155         fdt = kmalloc(sizeof(struct fdtable), GFP_KERNEL);
 156         if (!fdt)
 157                 goto out;
 158         fdt->max_fds = nr;
 159         data = alloc_fdmem(nr * sizeof(struct file *));
 160         if (!data)
 161                 goto out_fdt;
 162         fdt->fd = (struct file **)data;
 163         data = alloc_fdmem(max_t(unsigned int,
 164                                  2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
 165         if (!data)
 166                 goto out_arr;
 167         fdt->open_fds = (fd_set *)data;
 168         data += nr / BITS_PER_BYTE;
 169         fdt->close_on_exec = (fd_set *)data;
 170         INIT_RCU_HEAD(&fdt->rcu);
 171         fdt->next = NULL;
 172
 173         return fdt;
 174
 175 out_arr:
 176         free_fdarr(fdt);
 177 out_fdt:
 178         kfree(fdt);
 179 out:
 180         return NULL;
 181 }
 182
 183 /*
 184  * Expand the file descriptor table.
 185  * This function will allocate a new fdtable and both fd array and fdset, of
 186  * the given size.
 187  * Return <0 error code on error; 1 on successful completion.
 188  * The files->file_lock should be held on entry, and will be held on exit.
 189  */
 190 static int expand_fdtable(struct files_struct *files, int nr)
 191         __releases(files->file_lock)
 192         __acquires(files->file_lock)
 193 {
 194         struct fdtable *new_fdt, *cur_fdt;
 195
 196         spin_unlock(&files->file_lock);
 197         new_fdt = alloc_fdtable(nr);
 198         spin_lock(&files->file_lock);
 199         if (!new_fdt)
 200                 return -ENOMEM;
 201         /*
 202          * Check again since another task may have expanded the fd table while
 203          * we dropped the lock
 204          */
 205         cur_fdt = files_fdtable(files);
 206         if (nr >= cur_fdt->max_fds) {
 207                 /* Continue as planned */
 208                 copy_fdtable(new_fdt, cur_fdt);
 209                 rcu_assign_pointer(files->fdt, new_fdt);
 210                 if (cur_fdt->max_fds > NR_OPEN_DEFAULT)
 211                         free_fdtable(cur_fdt);
 212         } else {
 213                 /* Somebody else expanded, so undo our attempt */
 214                 free_fdarr(new_fdt);
 215                 free_fdset(new_fdt);
 216                 kfree(new_fdt);
 217         }
 218         return 1;
 219 }
 220
 221 /*
 222  * Expand files.
 223  * This function will expand the file structures, if the requested size exceeds
 224  * the current capacity and there is room for expansion.
 225  * Return <0 error code on error; 0 when nothing done; 1 when files were
 226  * expanded and execution may have blocked.
 227  * The files->file_lock should be held on entry, and will be held on exit.
 228  */
 229 int expand_files(struct files_struct *files, int nr)
 230 {
 231         struct fdtable *fdt;
 232
 233         fdt = files_fdtable(files);
 234         /* Do we need to expand? */
 235         if (nr < fdt->max_fds)
 236                 return 0;
 237         /* Can we expand? */
 238         if (nr >= sysctl_nr_open)
 239                 return -EMFILE;
 240
 241         /* All good, so we try */
 242         return expand_fdtable(files, nr);
 243 }
 244
 245 static void __devinit fdtable_defer_list_init(int cpu)
 246 {
 247         struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
 248         spin_lock_init(&fddef->lock);
 249         INIT_WORK(&fddef->wq, free_fdtable_work);
 250         fddef->next = NULL;
 251 }
 252
 253 void __init files_defer_init(void)
 254 {
 255         int i;
 256         for_each_possible_cpu(i)
 257                 fdtable_defer_list_init(i);
 258 }