| 1 |
// SPDX-License-Identifier: GPL-2.0 |
| 2 |
/* |
| 3 |
* Copyright 2017 Omnibond Systems, L.L.C. |
| 4 |
*/ |
| 5 |
|
| 6 |
#include <linux/filelock.h> |
| 7 |
#include "protocol.h" |
| 8 |
#include "orangefs-kernel.h" |
| 9 |
#include "orangefs-bufmap.h" |
| 10 |
|
| 11 |
struct orangefs_dir_part { |
| 12 |
struct orangefs_dir_part *next; |
| 13 |
size_t len; |
| 14 |
}; |
| 15 |
|
| 16 |
struct orangefs_dir { |
| 17 |
__u64 token; |
| 18 |
struct orangefs_dir_part *part; |
| 19 |
loff_t end; |
| 20 |
int error; |
| 21 |
}; |
| 22 |
|
| 23 |
#define PART_SHIFT (24) |
| 24 |
#define PART_SIZE (1<<24) |
| 25 |
#define PART_MASK (~(PART_SIZE - 1)) |
| 26 |
|
| 27 |
/* |
| 28 |
* There can be up to 512 directory entries. Each entry is encoded as |
| 29 |
* follows: |
| 30 |
* 4 bytes: string size (n) |
| 31 |
* n bytes: string |
| 32 |
* 1 byte: trailing zero |
| 33 |
* padding to 8 bytes |
| 34 |
* 16 bytes: khandle |
| 35 |
* padding to 8 bytes |
| 36 |
* |
| 37 |
* The trailer_buf starts with a struct orangefs_readdir_response_s |
| 38 |
* which must be skipped to get to the directory data. |
| 39 |
* |
| 40 |
* The data which is received from the userspace daemon is termed a |
| 41 |
* part and is stored in a linked list in case more than one part is |
| 42 |
* needed for a large directory. |
| 43 |
* |
| 44 |
* The position pointer (ctx->pos) encodes the part and offset on which |
| 45 |
* to begin reading at. Bits above PART_SHIFT encode the part and bits |
| 46 |
* below PART_SHIFT encode the offset. Parts are stored in a linked |
| 47 |
* list which grows as data is received from the server. The overhead |
| 48 |
* associated with managing the list is presumed to be small compared to |
| 49 |
* the overhead of communicating with the server. |
| 50 |
* |
| 51 |
* As data is received from the server, it is placed at the end of the |
| 52 |
* part list. Data is parsed from the current position as it is needed. |
| 53 |
* When data is determined to be corrupt, it is either because the |
| 54 |
* userspace component has sent back corrupt data or because the file |
| 55 |
* pointer has been moved to an invalid location. Since the two cannot |
| 56 |
* be differentiated, return EIO. |
| 57 |
* |
| 58 |
* Part zero is synthesized to contains `.' and `..'. Part one is the |
| 59 |
* first part of the part list. |
| 60 |
*/ |
| 61 |
|
| 62 |
static int do_readdir(struct orangefs_dir *od, struct inode *inode, |
| 63 |
struct orangefs_kernel_op_s *op) |
| 64 |
{ |
| 65 |
struct orangefs_inode_s *oi = ORANGEFS_I(inode); |
| 66 |
struct orangefs_readdir_response_s *resp; |
| 67 |
int bufi, r; |
| 68 |
|
| 69 |
/* |
| 70 |
* Despite the badly named field, readdir does not use shared |
| 71 |
* memory. However, there are a limited number of readdir |
| 72 |
* slots, which must be allocated here. This flag simply tells |
| 73 |
* the op scheduler to return the op here for retry. |
| 74 |
*/ |
| 75 |
op->uses_shared_memory = 1; |
| 76 |
op->upcall.req.readdir.refn = oi->refn; |
| 77 |
op->upcall.req.readdir.token = od->token; |
| 78 |
op->upcall.req.readdir.max_dirent_count = |
| 79 |
ORANGEFS_MAX_DIRENT_COUNT_READDIR; |
| 80 |
|
| 81 |
again: |
| 82 |
bufi = orangefs_readdir_index_get(); |
| 83 |
if (bufi < 0) { |
| 84 |
od->error = bufi; |
| 85 |
return bufi; |
| 86 |
} |
| 87 |
|
| 88 |
op->upcall.req.readdir.buf_index = bufi; |
| 89 |
|
| 90 |
r = service_operation(op, "orangefs_readdir", |
| 91 |
get_interruptible_flag(inode)); |
| 92 |
|
| 93 |
orangefs_readdir_index_put(bufi); |
| 94 |
|
| 95 |
if (op_state_purged(op)) { |
| 96 |
if (r == -EAGAIN) { |
| 97 |
vfree(op->downcall.trailer_buf); |
| 98 |
goto again; |
| 99 |
} else if (r == -EIO) { |
| 100 |
vfree(op->downcall.trailer_buf); |
| 101 |
od->error = r; |
| 102 |
return r; |
| 103 |
} |
| 104 |
} |
| 105 |
|
| 106 |
if (r < 0) { |
| 107 |
vfree(op->downcall.trailer_buf); |
| 108 |
od->error = r; |
| 109 |
return r; |
| 110 |
} else if (op->downcall.status) { |
| 111 |
vfree(op->downcall.trailer_buf); |
| 112 |
od->error = op->downcall.status; |
| 113 |
return op->downcall.status; |
| 114 |
} |
| 115 |
|
| 116 |
/* |
| 117 |
* The maximum size is size per entry times the 512 entries plus |
| 118 |
* the header. This is well under the limit. |
| 119 |
*/ |
| 120 |
if (op->downcall.trailer_size > PART_SIZE) { |
| 121 |
vfree(op->downcall.trailer_buf); |
| 122 |
od->error = -EIO; |
| 123 |
return -EIO; |
| 124 |
} |
| 125 |
|
| 126 |
resp = (struct orangefs_readdir_response_s *) |
| 127 |
op->downcall.trailer_buf; |
| 128 |
od->token = resp->token; |
| 129 |
return 0; |
| 130 |
} |
| 131 |
|
| 132 |
static int parse_readdir(struct orangefs_dir *od, |
| 133 |
struct orangefs_kernel_op_s *op) |
| 134 |
{ |
| 135 |
struct orangefs_dir_part *part, *new; |
| 136 |
size_t count; |
| 137 |
|
| 138 |
count = 1; |
| 139 |
part = od->part; |
| 140 |
while (part) { |
| 141 |
count++; |
| 142 |
if (part->next) |
| 143 |
part = part->next; |
| 144 |
else |
| 145 |
break; |
| 146 |
} |
| 147 |
|
| 148 |
new = (void *)op->downcall.trailer_buf; |
| 149 |
new->next = NULL; |
| 150 |
new->len = op->downcall.trailer_size - |
| 151 |
sizeof(struct orangefs_readdir_response_s); |
| 152 |
if (!od->part) |
| 153 |
od->part = new; |
| 154 |
else |
| 155 |
part->next = new; |
| 156 |
count++; |
| 157 |
od->end = count << PART_SHIFT; |
| 158 |
|
| 159 |
return 0; |
| 160 |
} |
| 161 |
|
| 162 |
static int orangefs_dir_more(struct orangefs_dir *od, struct inode *inode) |
| 163 |
{ |
| 164 |
struct orangefs_kernel_op_s *op; |
| 165 |
int r; |
| 166 |
|
| 167 |
op = op_alloc(ORANGEFS_VFS_OP_READDIR); |
| 168 |
if (!op) { |
| 169 |
od->error = -ENOMEM; |
| 170 |
return -ENOMEM; |
| 171 |
} |
| 172 |
r = do_readdir(od, inode, op); |
| 173 |
if (r) { |
| 174 |
od->error = r; |
| 175 |
goto out; |
| 176 |
} |
| 177 |
r = parse_readdir(od, op); |
| 178 |
if (r) { |
| 179 |
od->error = r; |
| 180 |
goto out; |
| 181 |
} |
| 182 |
|
| 183 |
od->error = 0; |
| 184 |
out: |
| 185 |
op_release(op); |
| 186 |
return od->error; |
| 187 |
} |
| 188 |
|
| 189 |
static int fill_from_part(struct orangefs_dir_part *part, |
| 190 |
struct dir_context *ctx) |
| 191 |
{ |
| 192 |
const int offset = sizeof(struct orangefs_readdir_response_s); |
| 193 |
struct orangefs_khandle *khandle; |
| 194 |
__u32 *len, padlen; |
| 195 |
loff_t i; |
| 196 |
char *s; |
| 197 |
i = ctx->pos & ~PART_MASK; |
| 198 |
|
| 199 |
/* The file offset from userspace is too large. */ |
| 200 |
if (i > part->len) |
| 201 |
return 1; |
| 202 |
|
| 203 |
/* |
| 204 |
* If the seek pointer is positioned just before an entry it |
| 205 |
* should find the next entry. |
| 206 |
*/ |
| 207 |
if (i % 8) |
| 208 |
i = i + (8 - i%8)%8; |
| 209 |
|
| 210 |
while (i < part->len) { |
| 211 |
if (part->len < i + sizeof *len) |
| 212 |
break; |
| 213 |
len = (void *)part + offset + i; |
| 214 |
/* |
| 215 |
* len is the size of the string itself. padlen is the |
| 216 |
* total size of the encoded string. |
| 217 |
*/ |
| 218 |
padlen = (sizeof *len + *len + 1) + |
| 219 |
(8 - (sizeof *len + *len + 1)%8)%8; |
| 220 |
if (part->len < i + padlen + sizeof *khandle) |
| 221 |
goto next; |
| 222 |
s = (void *)part + offset + i + sizeof *len; |
| 223 |
if (s[*len] != 0) |
| 224 |
goto next; |
| 225 |
khandle = (void *)part + offset + i + padlen; |
| 226 |
if (!dir_emit(ctx, s, *len, |
| 227 |
orangefs_khandle_to_ino(khandle), |
| 228 |
DT_UNKNOWN)) |
| 229 |
return 0; |
| 230 |
i += padlen + sizeof *khandle; |
| 231 |
i = i + (8 - i%8)%8; |
| 232 |
BUG_ON(i > part->len); |
| 233 |
ctx->pos = (ctx->pos & PART_MASK) | i; |
| 234 |
continue; |
| 235 |
next: |
| 236 |
i += 8; |
| 237 |
} |
| 238 |
return 1; |
| 239 |
} |
| 240 |
|
| 241 |
static int orangefs_dir_fill(struct orangefs_dir *od, struct dir_context *ctx) |
| 242 |
{ |
| 243 |
struct orangefs_dir_part *part; |
| 244 |
size_t count; |
| 245 |
|
| 246 |
count = ((ctx->pos & PART_MASK) >> PART_SHIFT) - 1; |
| 247 |
|
| 248 |
part = od->part; |
| 249 |
while (part->next && count) { |
| 250 |
count--; |
| 251 |
part = part->next; |
| 252 |
} |
| 253 |
/* This means the userspace file offset is invalid. */ |
| 254 |
if (count) { |
| 255 |
od->error = -EIO; |
| 256 |
return -EIO; |
| 257 |
} |
| 258 |
|
| 259 |
while (part && part->len) { |
| 260 |
int r; |
| 261 |
r = fill_from_part(part, ctx); |
| 262 |
if (r < 0) { |
| 263 |
od->error = r; |
| 264 |
return r; |
| 265 |
} else if (r == 0) { |
| 266 |
/* Userspace buffer is full. */ |
| 267 |
break; |
| 268 |
} else { |
| 269 |
/* |
| 270 |
* The part ran out of data. Move to the next |
| 271 |
* part. */ |
| 272 |
ctx->pos = (ctx->pos & PART_MASK) + |
| 273 |
(1 << PART_SHIFT); |
| 274 |
part = part->next; |
| 275 |
} |
| 276 |
} |
| 277 |
return 0; |
| 278 |
} |
| 279 |
|
| 280 |
static loff_t orangefs_dir_llseek(struct file *file, loff_t offset, |
| 281 |
int whence) |
| 282 |
{ |
| 283 |
struct orangefs_dir *od = file->private_data; |
| 284 |
/* |
| 285 |
* Delete the stored data so userspace sees new directory |
| 286 |
* entries. |
| 287 |
*/ |
| 288 |
if (!whence && offset < od->end) { |
| 289 |
struct orangefs_dir_part *part = od->part; |
| 290 |
while (part) { |
| 291 |
struct orangefs_dir_part *next = part->next; |
| 292 |
vfree(part); |
| 293 |
part = next; |
| 294 |
} |
| 295 |
od->token = ORANGEFS_ITERATE_START; |
| 296 |
od->part = NULL; |
| 297 |
od->end = 1 << PART_SHIFT; |
| 298 |
} |
| 299 |
return default_llseek(file, offset, whence); |
| 300 |
} |
| 301 |
|
| 302 |
static int orangefs_dir_iterate(struct file *file, |
| 303 |
struct dir_context *ctx) |
| 304 |
{ |
| 305 |
struct orangefs_dir *od = file->private_data; |
| 306 |
struct inode *inode = file_inode(file); |
| 307 |
int r; |
| 308 |
|
| 309 |
if (od->error) |
| 310 |
return od->error; |
| 311 |
|
| 312 |
if (ctx->pos == 0) { |
| 313 |
if (!dir_emit_dot(file, ctx)) |
| 314 |
return 0; |
| 315 |
ctx->pos++; |
| 316 |
} |
| 317 |
if (ctx->pos == 1) { |
| 318 |
if (!dir_emit_dotdot(file, ctx)) |
| 319 |
return 0; |
| 320 |
ctx->pos = 1 << PART_SHIFT; |
| 321 |
} |
| 322 |
|
| 323 |
/* |
| 324 |
* The seek position is in the first synthesized part but is not |
| 325 |
* valid. |
| 326 |
*/ |
| 327 |
if ((ctx->pos & PART_MASK) == 0) |
| 328 |
return -EIO; |
| 329 |
|
| 330 |
r = 0; |
| 331 |
|
| 332 |
/* |
| 333 |
* Must read more if the user has sought past what has been read |
| 334 |
* so far. Stop a user who has sought past the end. |
| 335 |
*/ |
| 336 |
while (od->token != ORANGEFS_ITERATE_END && |
| 337 |
ctx->pos > od->end) { |
| 338 |
r = orangefs_dir_more(od, inode); |
| 339 |
if (r) |
| 340 |
return r; |
| 341 |
} |
| 342 |
if (od->token == ORANGEFS_ITERATE_END && ctx->pos > od->end) |
| 343 |
return -EIO; |
| 344 |
|
| 345 |
/* Then try to fill if there's any left in the buffer. */ |
| 346 |
if (ctx->pos < od->end) { |
| 347 |
r = orangefs_dir_fill(od, ctx); |
| 348 |
if (r) |
| 349 |
return r; |
| 350 |
} |
| 351 |
|
| 352 |
/* Finally get some more and try to fill. */ |
| 353 |
if (od->token != ORANGEFS_ITERATE_END) { |
| 354 |
r = orangefs_dir_more(od, inode); |
| 355 |
if (r) |
| 356 |
return r; |
| 357 |
r = orangefs_dir_fill(od, ctx); |
| 358 |
} |
| 359 |
|
| 360 |
return r; |
| 361 |
} |
| 362 |
|
| 363 |
static int orangefs_dir_open(struct inode *inode, struct file *file) |
| 364 |
{ |
| 365 |
struct orangefs_dir *od; |
| 366 |
file->private_data = kmalloc_obj(struct orangefs_dir); |
| 367 |
if (!file->private_data) |
| 368 |
return -ENOMEM; |
| 369 |
od = file->private_data; |
| 370 |
od->token = ORANGEFS_ITERATE_START; |
| 371 |
od->part = NULL; |
| 372 |
od->end = 1 << PART_SHIFT; |
| 373 |
od->error = 0; |
| 374 |
return 0; |
| 375 |
} |
| 376 |
|
| 377 |
static int orangefs_dir_release(struct inode *inode, struct file *file) |
| 378 |
{ |
| 379 |
struct orangefs_dir *od = file->private_data; |
| 380 |
struct orangefs_dir_part *part = od->part; |
| 381 |
while (part) { |
| 382 |
struct orangefs_dir_part *next = part->next; |
| 383 |
vfree(part); |
| 384 |
part = next; |
| 385 |
} |
| 386 |
kfree(od); |
| 387 |
return 0; |
| 388 |
} |
| 389 |
|
| 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, |
| 396 |
.setlease = generic_setlease, |
| 397 |
}; |