1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright 2017 Omnibond Systems, L.L.C.
7 #include "orangefs-kernel.h"
8 #include "orangefs-bufmap.h"
10 struct orangefs_dir_part {
11 struct orangefs_dir_part *next;
17 struct orangefs_dir_part *part;
22 #define PART_SHIFT (24)
23 #define PART_SIZE (1<<24)
24 #define PART_MASK (~(PART_SIZE - 1))
27 * There can be up to 512 directory entries. Each entry is encoded as
29 * 4 bytes: string size (n)
31 * 1 byte: trailing zero
36 * The trailer_buf starts with a struct orangefs_readdir_response_s
37 * which must be skipped to get to the directory data.
39 * The data which is received from the userspace daemon is termed a
40 * part and is stored in a linked list in case more than one part is
41 * needed for a large directory.
43 * The position pointer (ctx->pos) encodes the part and offset on which
44 * to begin reading at. Bits above PART_SHIFT encode the part and bits
45 * below PART_SHIFT encode the offset. Parts are stored in a linked
46 * list which grows as data is received from the server. The overhead
47 * associated with managing the list is presumed to be small compared to
48 * the overhead of communicating with the server.
50 * As data is received from the server, it is placed at the end of the
51 * part list. Data is parsed from the current position as it is needed.
52 * When data is determined to be corrupt, it is either because the
53 * userspace component has sent back corrupt data or because the file
54 * pointer has been moved to an invalid location. Since the two cannot
55 * be differentiated, return EIO.
57 * Part zero is synthesized to contains `.' and `..'. Part one is the
58 * first part of the part list.
61 static int do_readdir(struct orangefs_dir *od, struct inode *inode,
62 struct orangefs_kernel_op_s *op)
64 struct orangefs_inode_s *oi = ORANGEFS_I(inode);
65 struct orangefs_readdir_response_s *resp;
69 * Despite the badly named field, readdir does not use shared
70 * memory. However, there are a limited number of readdir
71 * slots, which must be allocated here. This flag simply tells
72 * the op scheduler to return the op here for retry.
74 op->uses_shared_memory = 1;
75 op->upcall.req.readdir.refn = oi->refn;
76 op->upcall.req.readdir.token = od->token;
77 op->upcall.req.readdir.max_dirent_count =
78 ORANGEFS_MAX_DIRENT_COUNT_READDIR;
81 bufi = orangefs_readdir_index_get();
87 op->upcall.req.readdir.buf_index = bufi;
89 r = service_operation(op, "orangefs_readdir",
90 get_interruptible_flag(inode));
92 orangefs_readdir_index_put(bufi);
94 if (op_state_purged(op)) {
96 vfree(op->downcall.trailer_buf);
98 } else if (r == -EIO) {
99 vfree(op->downcall.trailer_buf);
106 vfree(op->downcall.trailer_buf);
109 } else if (op->downcall.status) {
110 vfree(op->downcall.trailer_buf);
111 od->error = op->downcall.status;
112 return op->downcall.status;
116 * The maximum size is size per entry times the 512 entries plus
117 * the header. This is well under the limit.
119 if (op->downcall.trailer_size > PART_SIZE) {
120 vfree(op->downcall.trailer_buf);
125 resp = (struct orangefs_readdir_response_s *)
126 op->downcall.trailer_buf;
127 od->token = resp->token;
131 static int parse_readdir(struct orangefs_dir *od,
132 struct orangefs_kernel_op_s *op)
134 struct orangefs_dir_part *part, *new;
147 new = (void *)op->downcall.trailer_buf;
149 new->len = op->downcall.trailer_size -
150 sizeof(struct orangefs_readdir_response_s);
156 od->end = count << PART_SHIFT;
161 static int orangefs_dir_more(struct orangefs_dir *od, struct inode *inode)
163 struct orangefs_kernel_op_s *op;
166 op = op_alloc(ORANGEFS_VFS_OP_READDIR);
171 r = do_readdir(od, inode, op);
176 r = parse_readdir(od, op);
188 static int fill_from_part(struct orangefs_dir_part *part,
189 struct dir_context *ctx)
191 const int offset = sizeof(struct orangefs_readdir_response_s);
192 struct orangefs_khandle *khandle;
196 i = ctx->pos & ~PART_MASK;
198 /* The file offset from userspace is too large. */
203 * If the seek pointer is positioned just before an entry it
204 * should find the next entry.
209 while (i < part->len) {
210 if (part->len < i + sizeof *len)
212 len = (void *)part + offset + i;
214 * len is the size of the string itself. padlen is the
215 * total size of the encoded string.
217 padlen = (sizeof *len + *len + 1) +
218 (8 - (sizeof *len + *len + 1)%8)%8;
219 if (part->len < i + padlen + sizeof *khandle)
221 s = (void *)part + offset + i + sizeof *len;
224 khandle = (void *)part + offset + i + padlen;
225 if (!dir_emit(ctx, s, *len,
226 orangefs_khandle_to_ino(khandle),
229 i += padlen + sizeof *khandle;
231 BUG_ON(i > part->len);
232 ctx->pos = (ctx->pos & PART_MASK) | i;
240 static int orangefs_dir_fill(struct orangefs_dir *od, struct dir_context *ctx)
242 struct orangefs_dir_part *part;
245 count = ((ctx->pos & PART_MASK) >> PART_SHIFT) - 1;
248 while (part->next && count) {
252 /* This means the userspace file offset is invalid. */
258 while (part && part->len) {
260 r = fill_from_part(part, ctx);
265 /* Userspace buffer is full. */
269 * The part ran out of data. Move to the next
271 ctx->pos = (ctx->pos & PART_MASK) +
279 static loff_t orangefs_dir_llseek(struct file *file, loff_t offset,
282 struct orangefs_dir *od = file->private_data;
284 * Delete the stored data so userspace sees new directory
287 if (!whence && offset < od->end) {
288 struct orangefs_dir_part *part = od->part;
290 struct orangefs_dir_part *next = part->next;
294 od->token = ORANGEFS_ITERATE_START;
296 od->end = 1 << PART_SHIFT;
298 return default_llseek(file, offset, whence);
301 static int orangefs_dir_iterate(struct file *file,
302 struct dir_context *ctx)
304 struct orangefs_dir *od = file->private_data;
305 struct inode *inode = file_inode(file);
312 if (!dir_emit_dot(file, ctx))
317 if (!dir_emit_dotdot(file, ctx))
319 ctx->pos = 1 << PART_SHIFT;
323 * The seek position is in the first synthesized part but is not
326 if ((ctx->pos & PART_MASK) == 0)
332 * Must read more if the user has sought past what has been read
333 * so far. Stop a user who has sought past the end.
335 while (od->token != ORANGEFS_ITERATE_END &&
336 ctx->pos > od->end) {
337 r = orangefs_dir_more(od, inode);
341 if (od->token == ORANGEFS_ITERATE_END && ctx->pos > od->end)
344 /* Then try to fill if there's any left in the buffer. */
345 if (ctx->pos < od->end) {
346 r = orangefs_dir_fill(od, ctx);
351 /* Finally get some more and try to fill. */
352 if (od->token != ORANGEFS_ITERATE_END) {
353 r = orangefs_dir_more(od, inode);
356 r = orangefs_dir_fill(od, ctx);
362 static int orangefs_dir_open(struct inode *inode, struct file *file)
364 struct orangefs_dir *od;
365 file->private_data = kmalloc(sizeof(struct orangefs_dir),
367 if (!file->private_data)
369 od = file->private_data;
370 od->token = ORANGEFS_ITERATE_START;
372 od->end = 1 << PART_SHIFT;
377 static int orangefs_dir_release(struct inode *inode, struct file *file)
379 struct orangefs_dir *od = file->private_data;
380 struct orangefs_dir_part *part = od->part;
382 struct orangefs_dir_part *next = part->next;
390 const struct file_operations orangefs_dir_operations = {
391 .llseek = orangefs_dir_llseek,
392 .read = generic_read_dir,
393 .iterate_shared = orangefs_dir_iterate,
394 .open = orangefs_dir_open,
395 .release = orangefs_dir_release