inotify: invalid mask should return a error number but not set it
[linux-2.6.git] / fs / file.c
index 4c6f0ea..ba3f605 100644 (file)
--- a/fs/file.c
+++ b/fs/file.c
@@ -6,9 +6,12 @@
  *  Manage the dynamic fd arrays in the process files_struct.
  */
 
+#include <linux/export.h>
 #include <linux/fs.h>
 #include <linux/mm.h>
+#include <linux/mmzone.h>
 #include <linux/time.h>
+#include <linux/sched.h>
 #include <linux/slab.h>
 #include <linux/vmalloc.h>
 #include <linux/file.h>
@@ -26,6 +29,8 @@ struct fdtable_defer {
 };
 
 int sysctl_nr_open __read_mostly = 1024*1024;
+int sysctl_nr_open_min = BITS_PER_LONG;
+int sysctl_nr_open_max = 1024 * 1024; /* raised later */
 
 /*
  * We use this list to defer free fdtables that have vmalloced
@@ -35,28 +40,30 @@ int sysctl_nr_open __read_mostly = 1024*1024;
  */
 static DEFINE_PER_CPU(struct fdtable_defer, fdtable_defer_list);
 
-static inline void * alloc_fdmem(unsigned int size)
+static void *alloc_fdmem(size_t size)
 {
-       if (size <= PAGE_SIZE)
-               return kmalloc(size, GFP_KERNEL);
-       else
-               return vmalloc(size);
+       /*
+        * Very large allocations can stress page reclaim, so fall back to
+        * vmalloc() if the allocation size will be considered "large" by the VM.
+        */
+       if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
+               void *data = kmalloc(size, GFP_KERNEL|__GFP_NOWARN);
+               if (data != NULL)
+                       return data;
+       }
+       return vmalloc(size);
 }
 
-static inline void free_fdarr(struct fdtable *fdt)
+static void free_fdmem(void *ptr)
 {
-       if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *)))
-               kfree(fdt->fd);
-       else
-               vfree(fdt->fd);
+       is_vmalloc_addr(ptr) ? vfree(ptr) : kfree(ptr);
 }
 
-static inline void free_fdset(struct fdtable *fdt)
+static void __free_fdtable(struct fdtable *fdt)
 {
-       if (fdt->max_fds <= (PAGE_SIZE * BITS_PER_BYTE / 2))
-               kfree(fdt->open_fds);
-       else
-               vfree(fdt->open_fds);
+       free_fdmem(fdt->fd);
+       free_fdmem(fdt->open_fds);
+       kfree(fdt);
 }
 
 static void free_fdtable_work(struct work_struct *work)
@@ -71,9 +78,8 @@ static void free_fdtable_work(struct work_struct *work)
        spin_unlock_bh(&f->lock);
        while(fdt) {
                struct fdtable *next = fdt->next;
-               vfree(fdt->fd);
-               free_fdset(fdt);
-               kfree(fdt);
+
+               __free_fdtable(fdt);
                fdt = next;
        }
 }
@@ -94,7 +100,7 @@ void free_fdtable_rcu(struct rcu_head *rcu)
                                container_of(fdt, struct files_struct, fdtab));
                return;
        }
-       if (fdt->max_fds <= (PAGE_SIZE / sizeof(struct file *))) {
+       if (!is_vmalloc_addr(fdt->fd) && !is_vmalloc_addr(fdt->open_fds)) {
                kfree(fdt->fd);
                kfree(fdt->open_fds);
                kfree(fdt);
@@ -119,8 +125,6 @@ static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
        unsigned int cpy, set;
 
        BUG_ON(nfdt->max_fds < ofdt->max_fds);
-       if (ofdt->max_fds == 0)
-               return;
 
        cpy = ofdt->max_fds * sizeof(struct file *);
        set = (nfdt->max_fds - ofdt->max_fds) * sizeof(struct file *);
@@ -138,7 +142,7 @@ static void copy_fdtable(struct fdtable *nfdt, struct fdtable *ofdt)
 static struct fdtable * alloc_fdtable(unsigned int nr)
 {
        struct fdtable *fdt;
-       char *data;
+       void *data;
 
        /*
         * Figure out how many fds we actually want to support in this fdtable.
@@ -168,21 +172,21 @@ static struct fdtable * alloc_fdtable(unsigned int nr)
        data = alloc_fdmem(nr * sizeof(struct file *));
        if (!data)
                goto out_fdt;
-       fdt->fd = (struct file **)data;
-       data = alloc_fdmem(max_t(unsigned int,
+       fdt->fd = data;
+
+       data = alloc_fdmem(max_t(size_t,
                                 2 * nr / BITS_PER_BYTE, L1_CACHE_BYTES));
        if (!data)
                goto out_arr;
-       fdt->open_fds = (fd_set *)data;
+       fdt->open_fds = data;
        data += nr / BITS_PER_BYTE;
-       fdt->close_on_exec = (fd_set *)data;
-       INIT_RCU_HEAD(&fdt->rcu);
+       fdt->close_on_exec = data;
        fdt->next = NULL;
 
        return fdt;
 
 out_arr:
-       free_fdarr(fdt);
+       free_fdmem(fdt->fd);
 out_fdt:
        kfree(fdt);
 out:
@@ -212,9 +216,7 @@ static int expand_fdtable(struct files_struct *files, int nr)
         * caller and alloc_fdtable().  Cheaper to catch it here...
         */
        if (unlikely(new_fdt->max_fds <= nr)) {
-               free_fdarr(new_fdt);
-               free_fdset(new_fdt);
-               kfree(new_fdt);
+               __free_fdtable(new_fdt);
                return -EMFILE;
        }
        /*
@@ -230,9 +232,7 @@ static int expand_fdtable(struct files_struct *files, int nr)
                        free_fdtable(cur_fdt);
        } else {
                /* Somebody else expanded, so undo our attempt */
-               free_fdarr(new_fdt);
-               free_fdset(new_fdt);
-               kfree(new_fdt);
+               __free_fdtable(new_fdt);
        }
        return 1;
 }
@@ -250,9 +250,18 @@ int expand_files(struct files_struct *files, int nr)
        struct fdtable *fdt;
 
        fdt = files_fdtable(files);
+
+       /*
+        * N.B. For clone tasks sharing a files structure, this test
+        * will limit the total number of files that can be opened.
+        */
+       if (nr >= rlimit(RLIMIT_NOFILE))
+               return -EMFILE;
+
        /* Do we need to expand? */
        if (nr < fdt->max_fds)
                return 0;
+
        /* Can we expand? */
        if (nr >= sysctl_nr_open)
                return -EMFILE;
@@ -261,6 +270,131 @@ int expand_files(struct files_struct *files, int nr)
        return expand_fdtable(files, nr);
 }
 
+static int count_open_files(struct fdtable *fdt)
+{
+       int size = fdt->max_fds;
+       int i;
+
+       /* Find the last open fd */
+       for (i = size / BITS_PER_LONG; i > 0; ) {
+               if (fdt->open_fds[--i])
+                       break;
+       }
+       i = (i + 1) * BITS_PER_LONG;
+       return i;
+}
+
+/*
+ * Allocate a new files structure and copy contents from the
+ * passed in files structure.
+ * errorp will be valid only when the returned files_struct is NULL.
+ */
+struct files_struct *dup_fd(struct files_struct *oldf, int *errorp)
+{
+       struct files_struct *newf;
+       struct file **old_fds, **new_fds;
+       int open_files, size, i;
+       struct fdtable *old_fdt, *new_fdt;
+
+       *errorp = -ENOMEM;
+       newf = kmem_cache_alloc(files_cachep, GFP_KERNEL);
+       if (!newf)
+               goto out;
+
+       atomic_set(&newf->count, 1);
+
+       spin_lock_init(&newf->file_lock);
+       newf->next_fd = 0;
+       new_fdt = &newf->fdtab;
+       new_fdt->max_fds = NR_OPEN_DEFAULT;
+       new_fdt->close_on_exec = newf->close_on_exec_init;
+       new_fdt->open_fds = newf->open_fds_init;
+       new_fdt->fd = &newf->fd_array[0];
+       new_fdt->next = NULL;
+
+       spin_lock(&oldf->file_lock);
+       old_fdt = files_fdtable(oldf);
+       open_files = count_open_files(old_fdt);
+
+       /*
+        * Check whether we need to allocate a larger fd array and fd set.
+        */
+       while (unlikely(open_files > new_fdt->max_fds)) {
+               spin_unlock(&oldf->file_lock);
+
+               if (new_fdt != &newf->fdtab)
+                       __free_fdtable(new_fdt);
+
+               new_fdt = alloc_fdtable(open_files - 1);
+               if (!new_fdt) {
+                       *errorp = -ENOMEM;
+                       goto out_release;
+               }
+
+               /* beyond sysctl_nr_open; nothing to do */
+               if (unlikely(new_fdt->max_fds < open_files)) {
+                       __free_fdtable(new_fdt);
+                       *errorp = -EMFILE;
+                       goto out_release;
+               }
+
+               /*
+                * Reacquire the oldf lock and a pointer to its fd table
+                * who knows it may have a new bigger fd table. We need
+                * the latest pointer.
+                */
+               spin_lock(&oldf->file_lock);
+               old_fdt = files_fdtable(oldf);
+               open_files = count_open_files(old_fdt);
+       }
+
+       old_fds = old_fdt->fd;
+       new_fds = new_fdt->fd;
+
+       memcpy(new_fdt->open_fds, old_fdt->open_fds, open_files / 8);
+       memcpy(new_fdt->close_on_exec, old_fdt->close_on_exec, open_files / 8);
+
+       for (i = open_files; i != 0; i--) {
+               struct file *f = *old_fds++;
+               if (f) {
+                       get_file(f);
+               } else {
+                       /*
+                        * The fd may be claimed in the fd bitmap but not yet
+                        * instantiated in the files array if a sibling thread
+                        * is partway through open().  So make sure that this
+                        * fd is available to the new process.
+                        */
+                       __clear_open_fd(open_files - i, new_fdt);
+               }
+               rcu_assign_pointer(*new_fds++, f);
+       }
+       spin_unlock(&oldf->file_lock);
+
+       /* compute the remainder to be cleared */
+       size = (new_fdt->max_fds - open_files) * sizeof(struct file *);
+
+       /* This is long word aligned thus could use a optimized version */
+       memset(new_fds, 0, size);
+
+       if (new_fdt->max_fds > open_files) {
+               int left = (new_fdt->max_fds - open_files) / 8;
+               int start = open_files / BITS_PER_LONG;
+
+               memset(&new_fdt->open_fds[start], 0, left);
+               memset(&new_fdt->close_on_exec[start], 0, left);
+       }
+
+       rcu_assign_pointer(newf->fdt, new_fdt);
+
+       return newf;
+
+out_release:
+       kmem_cache_free(files_cachep, newf);
+out:
+       return NULL;
+}
+
 static void __devinit fdtable_defer_list_init(int cpu)
 {
        struct fdtable_defer *fddef = &per_cpu(fdtable_defer_list, cpu);
@@ -274,4 +408,77 @@ void __init files_defer_init(void)
        int i;
        for_each_possible_cpu(i)
                fdtable_defer_list_init(i);
+       sysctl_nr_open_max = min((size_t)INT_MAX, ~(size_t)0/sizeof(void *)) &
+                            -BITS_PER_LONG;
+}
+
+struct files_struct init_files = {
+       .count          = ATOMIC_INIT(1),
+       .fdt            = &init_files.fdtab,
+       .fdtab          = {
+               .max_fds        = NR_OPEN_DEFAULT,
+               .fd             = &init_files.fd_array[0],
+               .close_on_exec  = init_files.close_on_exec_init,
+               .open_fds       = init_files.open_fds_init,
+       },
+       .file_lock      = __SPIN_LOCK_UNLOCKED(init_task.file_lock),
+};
+
+/*
+ * allocate a file descriptor, mark it busy.
+ */
+int alloc_fd(unsigned start, unsigned flags)
+{
+       struct files_struct *files = current->files;
+       unsigned int fd;
+       int error;
+       struct fdtable *fdt;
+
+       spin_lock(&files->file_lock);
+repeat:
+       fdt = files_fdtable(files);
+       fd = start;
+       if (fd < files->next_fd)
+               fd = files->next_fd;
+
+       if (fd < fdt->max_fds)
+               fd = find_next_zero_bit(fdt->open_fds, fdt->max_fds, fd);
+
+       error = expand_files(files, fd);
+       if (error < 0)
+               goto out;
+
+       /*
+        * If we needed to expand the fs array we
+        * might have blocked - try again.
+        */
+       if (error)
+               goto repeat;
+
+       if (start <= files->next_fd)
+               files->next_fd = fd + 1;
+
+       __set_open_fd(fd, fdt);
+       if (flags & O_CLOEXEC)
+               __set_close_on_exec(fd, fdt);
+       else
+               __clear_close_on_exec(fd, fdt);
+       error = fd;
+#if 1
+       /* Sanity check */
+       if (rcu_dereference_raw(fdt->fd[fd]) != NULL) {
+               printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
+               rcu_assign_pointer(fdt->fd[fd], NULL);
+       }
+#endif
+
+out:
+       spin_unlock(&files->file_lock);
+       return error;
+}
+
+int get_unused_fd(void)
+{
+       return alloc_fd(0, 0);
 }
+EXPORT_SYMBOL(get_unused_fd);