]> git.baikalelectronics.ru Git - kernel.git/commitdiff
orangefs: rewrite readdir to fix several bugs
authorMartin Brandenburg <martin@omnibond.com>
Tue, 25 Apr 2017 19:37:59 +0000 (15:37 -0400)
committerMike Marshall <hubcap@omnibond.com>
Wed, 26 Apr 2017 18:33:00 +0000 (14:33 -0400)
In the past, readdir assumed that the user buffer will be large enough
that all entries from the server will fit.  If this was not true,
entries would be skipped.

Since it works now, request 512 entries rather than 96 per server
operation.

Signed-off-by: Martin Brandenburg <martin@omnibond.com>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
fs/orangefs/dir.c
fs/orangefs/downcall.h
fs/orangefs/orangefs-dev-proto.h

index 284373a57a082fa0e38422afbc49fa2e8519b0df..cf0ebb06b84e9e1c09091834fe9cff6375f2840f 100644 (file)
@@ -1,7 +1,5 @@
 /*
- * (C) 2001 Clemson University and The University of Chicago
- *
- * See COPYING in top-level directory.
+ * Copyright 2017 Omnibond Systems, L.L.C.
  */
 
 #include "protocol.h"
 #include "orangefs-bufmap.h"
 
 /*
- * decode routine used by kmod to deal with the blob sent from
- * userspace for readdirs. The blob contains zero or more of these
- * sub-blobs:
- *   __u32 - represents length of the character string that follows.
- *   string - between 1 and ORANGEFS_NAME_MAX bytes long.
- *   padding - (if needed) to cause the __u32 plus the string to be
- *             eight byte aligned.
- *   khandle - sizeof(khandle) bytes.
+ * There can be up to 512 directory entries.  Each entry is encoded as
+ * follows:
+ * 4 bytes: string size (n)
+ * n bytes: string
+ * 1 byte: trailing zero
+ * padding to 8 bytes
+ * 16 bytes: khandle
+ * padding to 8 bytes
  */
-static long decode_dirents(char *ptr, size_t size,
-                           struct orangefs_readdir_response_s *readdir)
-{
-       int i;
-       struct orangefs_readdir_response_s *rd =
-               (struct orangefs_readdir_response_s *) ptr;
-       char *buf = ptr;
-       int khandle_size = sizeof(struct orangefs_khandle);
-       size_t offset = offsetof(struct orangefs_readdir_response_s,
-                               dirent_array);
-       /* 8 reflects eight byte alignment */
-       int smallest_blob = khandle_size + 8;
-       __u32 len;
-       int aligned_len;
-       int sizeof_u32 = sizeof(__u32);
-       long ret;
-
-       gossip_debug(GOSSIP_DIR_DEBUG, "%s: size:%zu:\n", __func__, size);
+#define MAX_DIRECTORY ((4 + 257 + 3 + 16)*512)
 
-       /* size is = offset on empty dirs, > offset on non-empty dirs... */
-       if (size < offset) {
-               gossip_err("%s: size:%zu: offset:%zu:\n",
-                          __func__,
-                          size,
-                          offset);
-               ret = -EINVAL;
-               goto out;
-       }
-
-        if ((size == offset) && (readdir->orangefs_dirent_outcount != 0)) {
-               gossip_err("%s: size:%zu: dirent_outcount:%d:\n",
-                          __func__,
-                          size,
-                          readdir->orangefs_dirent_outcount);
-               ret = -EINVAL;
-               goto out;
-       }
-
-       readdir->token = rd->token;
-       readdir->orangefs_dirent_outcount = rd->orangefs_dirent_outcount;
-       readdir->dirent_array = kcalloc(readdir->orangefs_dirent_outcount,
-                                       sizeof(*readdir->dirent_array),
-                                       GFP_KERNEL);
-       if (readdir->dirent_array == NULL) {
-               gossip_err("%s: kcalloc failed.\n", __func__);
-               ret = -ENOMEM;
-               goto out;
-       }
-
-       buf += offset;
-       size -= offset;
-
-       for (i = 0; i < readdir->orangefs_dirent_outcount; i++) {
-               if (size < smallest_blob) {
-                       gossip_err("%s: size:%zu: smallest_blob:%d:\n",
-                                  __func__,
-                                  size,
-                                  smallest_blob);
-                       ret = -EINVAL;
-                       goto free;
-               }
-
-               len = *(__u32 *)buf;
-               if ((len < 1) || (len > ORANGEFS_NAME_MAX)) {
-                       gossip_err("%s: len:%d:\n", __func__, len);
-                       ret = -EINVAL;
-                       goto free;
-               }
-
-               gossip_debug(GOSSIP_DIR_DEBUG,
-                            "%s: size:%zu: len:%d:\n",
-                            __func__,
-                            size,
-                            len);
-
-               readdir->dirent_array[i].d_name = buf + sizeof_u32;
-               readdir->dirent_array[i].d_length = len;
-
-               /*
-                * Calculate "aligned" length of this string and its
-                * associated __u32 descriptor.
-                */
-               aligned_len = ((sizeof_u32 + len + 1) + 7) & ~7;
-               gossip_debug(GOSSIP_DIR_DEBUG,
-                            "%s: aligned_len:%d:\n",
-                            __func__,
-                            aligned_len);
-
-               /*
-                * The end of the blob should coincide with the end
-                * of the last sub-blob.
-                */
-               if (size < aligned_len + khandle_size) {
-                       gossip_err("%s: ran off the end of the blob.\n",
-                                  __func__);
-                       ret = -EINVAL;
-                       goto free;
-               }
-               size -= aligned_len + khandle_size;
-
-               buf += aligned_len;
-
-               readdir->dirent_array[i].khandle =
-                       *(struct orangefs_khandle *) buf;
-               buf += khandle_size;
-       }
-       ret = buf - ptr;
-       gossip_debug(GOSSIP_DIR_DEBUG, "%s: returning:%ld:\n", __func__, ret);
-       goto out;
-
-free:
-       kfree(readdir->dirent_array);
-       readdir->dirent_array = NULL;
-
-out:
-       return ret;
-}
+struct orangefs_dir {
+       __u64 token;
+       void *directory;
+       size_t i, len;
+       int error;
+};
 
 /*
- * Read directory entries from an instance of an open directory.
+ * The userspace component sends several directory entries of the
+ * following format.  The first four bytes are the string length not
+ * including a trailing zero byte.  This is followed by the string and a
+ * trailing zero padded to the next four byte boundry.  This is followed
+ * by the sixteen byte khandle padded to the next eight byte boundry.
+ *
+ * The trailer_buf starts with a struct orangefs_readdir_response_s
+ * which must be skipped to get to the directory data.
  */
-static int orangefs_readdir(struct file *file, struct dir_context *ctx)
-{
-       int ret = 0;
-       int buffer_index;
-       /*
-        * ptoken supports Orangefs' distributed directory logic, added
-        * in 2.9.2.
-        */
-       __u64 *ptoken = file->private_data;
-       __u64 pos = 0;
-       ino_t ino = 0;
-       struct dentry *dentry = file->f_path.dentry;
-       struct orangefs_kernel_op_s *new_op = NULL;
-       struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(dentry->d_inode);
-       struct orangefs_readdir_response_s readdir_response;
-       void *dents_buf;
-       int i = 0;
-       int len = 0;
-       ino_t current_ino = 0;
-       char *current_entry = NULL;
-       long bytes_decoded;
-
-       gossip_debug(GOSSIP_DIR_DEBUG,
-                    "%s: ctx->pos:%lld, ptoken = %llu\n",
-                    __func__,
-                    lld(ctx->pos),
-                    llu(*ptoken));
-
-       pos = (__u64) ctx->pos;
-
-       /* are we done? */
-       if (pos == ORANGEFS_READDIR_END) {
-               gossip_debug(GOSSIP_DIR_DEBUG,
-                            "Skipping to termination path\n");
-               return 0;
-       }
-
-       gossip_debug(GOSSIP_DIR_DEBUG,
-                    "orangefs_readdir called on %pd (pos=%llu)\n",
-                    dentry, llu(pos));
-
-       memset(&readdir_response, 0, sizeof(readdir_response));
 
-       new_op = op_alloc(ORANGEFS_VFS_OP_READDIR);
-       if (!new_op)
+static int orangefs_dir_more(struct orangefs_inode_s *oi,
+    struct orangefs_dir *od, struct dentry *dentry)
+{
+       const size_t offset =
+           sizeof(struct orangefs_readdir_response_s);
+       struct orangefs_readdir_response_s *resp;
+       struct orangefs_kernel_op_s *op;
+       int bufi, r;
+
+       op = op_alloc(ORANGEFS_VFS_OP_READDIR);
+       if (!op) {
+               od->error = -ENOMEM;
                return -ENOMEM;
+       }
 
        /*
-        * Only the indices are shared. No memory is actually shared, but the
-        * mechanism is used.
+        * Despite the badly named field, readdir does not use shared
+        * memory.  However, there are a limited number of readdir
+        * slots, which must be allocated here.  This flag simply tells
+        * the op scheduler to return the op here for retry.
         */
-       new_op->uses_shared_memory = 1;
-       new_op->upcall.req.readdir.refn = orangefs_inode->refn;
-       new_op->upcall.req.readdir.max_dirent_count =
+       op->uses_shared_memory = 1;
+       op->upcall.req.readdir.refn = oi->refn;
+       op->upcall.req.readdir.token = od->token;
+       op->upcall.req.readdir.max_dirent_count =
            ORANGEFS_MAX_DIRENT_COUNT_READDIR;
 
-       gossip_debug(GOSSIP_DIR_DEBUG,
-                    "%s: upcall.req.readdir.refn.khandle: %pU\n",
-                    __func__,
-                    &new_op->upcall.req.readdir.refn.khandle);
-
-       new_op->upcall.req.readdir.token = *ptoken;
-
-get_new_buffer_index:
-       buffer_index = orangefs_readdir_index_get();
-       if (buffer_index < 0) {
-               ret = buffer_index;
-               gossip_lerr("orangefs_readdir: orangefs_readdir_index_get() failure (%d)\n",
-                           ret);
-               goto out_free_op;
+again:
+       bufi = orangefs_readdir_index_get();
+       if (bufi < 0) {
+               op_release(op);
+               od->error = bufi;
+               return bufi;
        }
-       new_op->upcall.req.readdir.buf_index = buffer_index;
 
-       ret = service_operation(new_op,
-                               "orangefs_readdir",
-                               get_interruptible_flag(dentry->d_inode));
+       op->upcall.req.readdir.buf_index = bufi;
 
-       gossip_debug(GOSSIP_DIR_DEBUG,
-                    "Readdir downcall status is %d.  ret:%d\n",
-                    new_op->downcall.status,
-                    ret);
+       r = service_operation(op, "orangefs_readdir",
+           get_interruptible_flag(dentry->d_inode));
 
-       orangefs_readdir_index_put(buffer_index);
+       orangefs_readdir_index_put(bufi);
 
-       if (ret == -EAGAIN && op_state_purged(new_op)) {
-               /* Client-core indices are invalid after it restarted. */
-               gossip_debug(GOSSIP_DIR_DEBUG,
-                       "%s: Getting new buffer_index for retry of readdir..\n",
-                        __func__);
-               goto get_new_buffer_index;
-       }
-
-       if (ret == -EIO && op_state_purged(new_op)) {
-               gossip_err("%s: Client is down. Aborting readdir call.\n",
-                       __func__);
-               goto out_free_op;
+       if (op_state_purged(op)) {
+               if (r == -EAGAIN) {
+                       vfree(op->downcall.trailer_buf);
+                       goto again;
+               } else if (r == -EIO) {
+                       vfree(op->downcall.trailer_buf);
+                       op_release(op);
+                       od->error = r;
+                       return r;
+               }
        }
 
-       if (ret < 0 || new_op->downcall.status != 0) {
-               gossip_debug(GOSSIP_DIR_DEBUG,
-                            "Readdir request failed.  Status:%d\n",
-                            new_op->downcall.status);
-               if (ret >= 0)
-                       ret = new_op->downcall.status;
-               goto out_free_op;
-       }
+       if (r < 0) {
+               vfree(op->downcall.trailer_buf);
+               op_release(op);
+               od->error = r;
+               return r;
+       } else if (op->downcall.status) {
+               vfree(op->downcall.trailer_buf);
+               op_release(op);
+               od->error = op->downcall.status;
+               return op->downcall.status;
+       }
+
+       resp = (struct orangefs_readdir_response_s *)
+           op->downcall.trailer_buf;
+       od->token = resp->token;
+
+       if (od->len + op->downcall.trailer_size - offset <=
+           MAX_DIRECTORY) {
+               memcpy(od->directory + od->len,
+                   op->downcall.trailer_buf + offset,
+                   op->downcall.trailer_size - offset);
+               od->len += op->downcall.trailer_size - offset;
+       } else {
+               /* This limit was chosen based on protocol limits. */
+               gossip_err("orangefs_dir_more: userspace sent too much data\n");
+               vfree(op->downcall.trailer_buf);
+               op_release(op);
+               od->error = -EIO;
+               return -EIO;
+       }
+
+       vfree(op->downcall.trailer_buf);
+       op_release(op);
+       return 0;
+}
 
-       dents_buf = new_op->downcall.trailer_buf;
-       if (dents_buf == NULL) {
-               gossip_err("Invalid NULL buffer in readdir response\n");
-               ret = -ENOMEM;
-               goto out_free_op;
-       }
+static int orangefs_dir_fill(struct orangefs_inode_s *oi,
+    struct orangefs_dir *od, struct dentry *dentry,
+    struct dir_context *ctx)
+{
+       struct orangefs_khandle *khandle;
+       __u32 *len, padlen;
+       char *s;
+       while (od->i < od->len) {
+               if (od->len < od->i + sizeof *len)
+                       goto eio;
+               len = od->directory + od->i;
+               /*
+                * len is the size of the string itself.  padlen is the
+                * total size of the encoded string.
+                */
+               padlen = (sizeof *len + *len + 1) +
+                   (4 - (sizeof *len + *len + 1)%8)%8;
+               if (od->len < od->i + padlen + sizeof *khandle)
+                       goto eio;
+               s = od->directory + od->i + sizeof *len;
+               if (s[*len] != 0)
+                       goto eio;
+               khandle = od->directory + od->i + padlen;
+
+               if (!dir_emit(ctx, s, *len,
+                   orangefs_khandle_to_ino(khandle), DT_UNKNOWN))
+                       return 0;
+               od->i += padlen + sizeof *khandle;
+               od->i = od->i + (8 - od->i%8)%8;
+               ctx->pos = 2 + od->i;
+       }
+       BUG_ON(od->i > od->len);
+       return 0;
+eio:
+       gossip_err("orangefs_dir_fill: userspace returns corrupt data\n");
+       od->error = -EIO;
+       return -EIO;
+}
 
-       bytes_decoded = decode_dirents(dents_buf, new_op->downcall.trailer_size,
-                                       &readdir_response);
-       if (bytes_decoded < 0) {
-               ret = bytes_decoded;
-               gossip_err("Could not decode readdir from buffer %d\n", ret);
-               goto out_vfree;
-       }
+static int orangefs_dir_iterate(struct file *file,
+    struct dir_context *ctx)
+{
+       struct orangefs_inode_s *oi;
+       struct orangefs_dir *od;
+       struct dentry *dentry;
+       int r;
 
-       if (bytes_decoded != new_op->downcall.trailer_size) {
-               gossip_err("orangefs_readdir: # bytes decoded (%ld) "
-                          "!= trailer size (%ld)\n",
-                          bytes_decoded,
-                          (long)new_op->downcall.trailer_size);
-               ret = -EINVAL;
-               goto out_destroy_handle;
-       }
+       dentry = file->f_path.dentry;
+       oi = ORANGEFS_I(dentry->d_inode);
+       od = file->private_data;
 
-       /*
-        *  orangefs doesn't actually store dot and dot-dot, but
-        *  we need to have them represented.
-        */
-       if (pos == 0) {
-               ino = get_ino_from_khandle(dentry->d_inode);
-               gossip_debug(GOSSIP_DIR_DEBUG,
-                            "%s: calling dir_emit of \".\" with pos = %llu\n",
-                            __func__,
-                            llu(pos));
-               ret = dir_emit(ctx, ".", 1, ino, DT_DIR);
-               pos += 1;
-       }
+       if (od->error)
+               return od->error;
 
-       if (pos == 1) {
-               ino = get_parent_ino_from_dentry(dentry);
-               gossip_debug(GOSSIP_DIR_DEBUG,
-                            "%s: calling dir_emit of \"..\" with pos = %llu\n",
-                            __func__,
-                            llu(pos));
-               ret = dir_emit(ctx, "..", 2, ino, DT_DIR);
-               pos += 1;
+       if (ctx->pos == 0) {
+               if (!dir_emit_dot(file, ctx))
+                       return 0;
+               ctx->pos++;
        }
-
-       /*
-        * we stored ORANGEFS_ITERATE_NEXT in ctx->pos last time around
-        * to prevent "finding" dot and dot-dot on any iteration
-        * other than the first.
-        */
-       if (ctx->pos == ORANGEFS_ITERATE_NEXT)
-               ctx->pos = 0;
-
-       gossip_debug(GOSSIP_DIR_DEBUG,
-                    "%s: dirent_outcount:%d:\n",
-                    __func__,
-                    readdir_response.orangefs_dirent_outcount);
-       for (i = ctx->pos;
-            i < readdir_response.orangefs_dirent_outcount;
-            i++) {
-               len = readdir_response.dirent_array[i].d_length;
-               current_entry = readdir_response.dirent_array[i].d_name;
-               current_ino = orangefs_khandle_to_ino(
-                       &readdir_response.dirent_array[i].khandle);
-
-               gossip_debug(GOSSIP_DIR_DEBUG,
-                            "calling dir_emit for %s with len %d"
-                            ", ctx->pos %ld\n",
-                            current_entry,
-                            len,
-                            (unsigned long)ctx->pos);
-               /*
-                * type is unknown. We don't return object type
-                * in the dirent_array. This leaves getdents
-                * clueless about type.
-                */
-               ret =
-                   dir_emit(ctx, current_entry, len, current_ino, DT_UNKNOWN);
-               if (!ret)
-                       break;
+       if (ctx->pos == 1) {
+               if (!dir_emit_dotdot(file, ctx))
+                       return 0;
                ctx->pos++;
-               gossip_debug(GOSSIP_DIR_DEBUG,
-                             "%s: ctx->pos:%lld\n",
-                             __func__,
-                             lld(ctx->pos));
-
        }
 
-       /*
-        * we ran all the way through the last batch, set up for
-        * getting another batch...
-        */
-       if (ret) {
-               *ptoken = readdir_response.token;
-               ctx->pos = ORANGEFS_ITERATE_NEXT;
+       r = 0;
+
+       if (od->i < od->len) {
+               r = orangefs_dir_fill(oi, od, dentry, ctx);
+               if (r)
+                       return r;
        }
 
-       /*
-        * Did we hit the end of the directory?
-        */
-       if (readdir_response.token == ORANGEFS_READDIR_END) {
-               gossip_debug(GOSSIP_DIR_DEBUG,
-               "End of dir detected; setting ctx->pos to ORANGEFS_READDIR_END.\n");
-               ctx->pos = ORANGEFS_READDIR_END;
+       if (od->token != ORANGEFS_READDIR_END) {
+               r = orangefs_dir_more(oi, od, dentry);
+               if (r)
+                       return r;
+               r = orangefs_dir_fill(oi, od, dentry, ctx);
        }
 
-out_destroy_handle:
-       /* kfree(NULL) is safe */
-       kfree(readdir_response.dirent_array);
-out_vfree:
-       gossip_debug(GOSSIP_DIR_DEBUG, "vfree %p\n", dents_buf);
-       vfree(dents_buf);
-out_free_op:
-       op_release(new_op);
-       gossip_debug(GOSSIP_DIR_DEBUG, "orangefs_readdir returning %d\n", ret);
-       return ret;
+       return r;
 }
 
 static int orangefs_dir_open(struct inode *inode, struct file *file)
 {
-       __u64 *ptoken;
-
-       file->private_data = kmalloc(sizeof(__u64), GFP_KERNEL);
+       struct orangefs_dir *od;
+       file->private_data = kmalloc(sizeof(struct orangefs_dir),
+           GFP_KERNEL);
        if (!file->private_data)
                return -ENOMEM;
-
-       ptoken = file->private_data;
-       *ptoken = ORANGEFS_READDIR_START;
+       od = file->private_data;
+       od->token = ORANGEFS_READDIR_START;
+       /*
+        * XXX: It seems wasteful to allocate such a large buffer for
+        * each request.  Most will be much smaller.
+        */
+       od->directory = alloc_pages_exact(MAX_DIRECTORY, GFP_KERNEL);
+       if (!od->directory) {
+               kfree(file->private_data);
+               return -ENOMEM;
+       }
+       od->i = 0;
+       od->len = 0;
+       od->error = 0;
        return 0;
 }
 
 static int orangefs_dir_release(struct inode *inode, struct file *file)
 {
+       struct orangefs_dir *od = file->private_data;
        orangefs_flush_inode(inode);
-       kfree(file->private_data);
+       free_pages_exact(od->directory, MAX_DIRECTORY);
+       kfree(od);
        return 0;
 }
 
-/** ORANGEFS implementation of VFS directory operations */
 const struct file_operations orangefs_dir_operations = {
        .read = generic_read_dir,
-       .iterate = orangefs_readdir,
+       .iterate = orangefs_dir_iterate,
        .open = orangefs_dir_open,
-       .release = orangefs_dir_release,
+       .release = orangefs_dir_release
 };
index 3b8923f8bf21d6902f0d8f440e2102cb57be94bf..163001c9550184feff1255364a3b59a61de1ecf9 100644 (file)
@@ -40,16 +40,6 @@ struct orangefs_mkdir_response {
        struct orangefs_object_kref refn;
 };
 
-/*
- * duplication of some system interface structures so that I don't have
- * to allocate extra memory
- */
-struct orangefs_dirent {
-       char *d_name;
-       int d_length;
-       struct orangefs_khandle khandle;
-};
-
 struct orangefs_statfs_response {
        __s64 block_size;
        __s64 blocks_total;
@@ -131,12 +121,16 @@ struct orangefs_downcall_s {
        } resp;
 };
 
+/*
+ * The readdir response comes in the trailer.  It is followed by the
+ * directory entries as described in dir.c.
+ */
+
 struct orangefs_readdir_response_s {
        __u64 token;
        __u64 directory_version;
        __u32 __pad2;
        __u32 orangefs_dirent_outcount;
-       struct orangefs_dirent *dirent_array;
 };
 
 #endif /* __DOWNCALL_H */
index f380f9ed1b286a7eedfb41fa5f7f8f9259abfdf3..efe08c763e566bffc500439645757966c65cc93c 100644 (file)
  */
 #define ORANGEFS_MAX_DEBUG_STRING_LEN  0x00000800
 
-/*
- * The maximum number of directory entries in a single request is 96.
- * XXX: Why can this not be higher. The client-side code can handle up to 512.
- * XXX: What happens if we expect more than the client can return?
- */
-#define ORANGEFS_MAX_DIRENT_COUNT_READDIR 96
+#define ORANGEFS_MAX_DIRENT_COUNT_READDIR 512
 
 #include "upcall.h"
 #include "downcall.h"