kernel/bpf/task_iter.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (c) 2020 Facebook */
   3
   4 #include <linux/init.h>
   5 #include <linux/namei.h>
   6 #include <linux/pid_namespace.h>
   7 #include <linux/fs.h>
   8 #include <linux/fdtable.h>
   9 #include <linux/filter.h>
  10 #include <linux/bpf_mem_alloc.h>
  11 #include <linux/btf_ids.h>
  12 #include <linux/mm_types.h>
  13 #include "mmap_unlock_work.h"
  14
  15 static const char * const iter_task_type_names[] = {
  16         "ALL",
  17         "TID",
  18         "PID",
  19 };
  20
  21 struct bpf_iter_seq_task_common {
  22         struct pid_namespace *ns;
  23         enum bpf_iter_task_type type;
  24         u32 pid;
  25         u32 pid_visiting;
  26 };
  27
  28 struct bpf_iter_seq_task_info {
  29         /* The first field must be struct bpf_iter_seq_task_common.
  30          * this is assumed by {init, fini}_seq_pidns() callback functions.
  31          */
  32         struct bpf_iter_seq_task_common common;
  33         u32 tid;
  34 };
  35
  36 static struct task_struct *task_group_seq_get_next(struct bpf_iter_seq_task_common *common,
  37                                                    u32 *tid,
  38                                                    bool skip_if_dup_files)
  39 {
  40         struct task_struct *task;
  41         struct pid *pid;
  42         u32 next_tid;
  43
  44         if (!*tid) {
  45                 /* The first time, the iterator calls this function. */
  46                 pid = find_pid_ns(common->pid, common->ns);
  47                 task = get_pid_task(pid, PIDTYPE_TGID);
  48                 if (!task)
  49                         return NULL;
  50
  51                 *tid = common->pid;
  52                 common->pid_visiting = common->pid;
  53
  54                 return task;
  55         }
  56
  57         /* If the control returns to user space and comes back to the
  58          * kernel again, *tid and common->pid_visiting should be the
  59          * same for task_seq_start() to pick up the correct task.
  60          */
  61         if (*tid == common->pid_visiting) {
  62                 pid = find_pid_ns(common->pid_visiting, common->ns);
  63                 task = get_pid_task(pid, PIDTYPE_PID);
  64
  65                 return task;
  66         }
  67
  68         task = find_task_by_pid_ns(common->pid_visiting, common->ns);
  69         if (!task)
  70                 return NULL;
  71
  72 retry:
  73         task = __next_thread(task);
  74         if (!task)
  75                 return NULL;
  76
  77         next_tid = __task_pid_nr_ns(task, PIDTYPE_PID, common->ns);
  78         if (!next_tid)
  79                 goto retry;
  80
  81         if (skip_if_dup_files && task->files == task->group_leader->files)
  82                 goto retry;
  83
  84         *tid = common->pid_visiting = next_tid;
  85         get_task_struct(task);
  86         return task;
  87 }
  88
  89 static struct task_struct *task_seq_get_next(struct bpf_iter_seq_task_common *common,
  90                                              u32 *tid,
  91                                              bool skip_if_dup_files)
  92 {
  93         struct task_struct *task = NULL;
  94         struct pid *pid;
  95
  96         if (common->type == BPF_TASK_ITER_TID) {
  97                 if (*tid && *tid != common->pid)
  98                         return NULL;
  99                 rcu_read_lock();
 100                 pid = find_pid_ns(common->pid, common->ns);
 101                 if (pid) {
 102                         task = get_pid_task(pid, PIDTYPE_TGID);
 103                         *tid = common->pid;
 104                 }
 105                 rcu_read_unlock();
 106
 107                 return task;
 108         }
 109
 110         if (common->type == BPF_TASK_ITER_TGID) {
 111                 rcu_read_lock();
 112                 task = task_group_seq_get_next(common, tid, skip_if_dup_files);
 113                 rcu_read_unlock();
 114
 115                 return task;
 116         }
 117
 118         rcu_read_lock();
 119 retry:
 120         pid = find_ge_pid(*tid, common->ns);
 121         if (pid) {
 122                 *tid = pid_nr_ns(pid, common->ns);
 123                 task = get_pid_task(pid, PIDTYPE_PID);
 124                 if (!task) {
 125                         ++*tid;
 126                         goto retry;
 127                 } else if (skip_if_dup_files && !thread_group_leader(task) &&
 128                            task->files == task->group_leader->files) {
 129                         put_task_struct(task);
 130                         task = NULL;
 131                         ++*tid;
 132                         goto retry;
 133                 }
 134         }
 135         rcu_read_unlock();
 136
 137         return task;
 138 }
 139
 140 static void *task_seq_start(struct seq_file *seq, loff_t *pos)
 141 {
 142         struct bpf_iter_seq_task_info *info = seq->private;
 143         struct task_struct *task;
 144
 145         task = task_seq_get_next(&info->common, &info->tid, false);
 146         if (!task)
 147                 return NULL;
 148
 149         if (*pos == 0)
 150                 ++*pos;
 151         return task;
 152 }
 153
 154 static void *task_seq_next(struct seq_file *seq, void *v, loff_t *pos)
 155 {
 156         struct bpf_iter_seq_task_info *info = seq->private;
 157         struct task_struct *task;
 158
 159         ++*pos;
 160         ++info->tid;
 161         put_task_struct((struct task_struct *)v);
 162         task = task_seq_get_next(&info->common, &info->tid, false);
 163         if (!task)
 164                 return NULL;
 165
 166         return task;
 167 }
 168
 169 struct bpf_iter__task {
 170         __bpf_md_ptr(struct bpf_iter_meta *, meta);
 171         __bpf_md_ptr(struct task_struct *, task);
 172 };
 173
 174 DEFINE_BPF_ITER_FUNC(task, struct bpf_iter_meta *meta, struct task_struct *task)
 175
 176 static int __task_seq_show(struct seq_file *seq, struct task_struct *task,
 177                            bool in_stop)
 178 {
 179         struct bpf_iter_meta meta;
 180         struct bpf_iter__task ctx;
 181         struct bpf_prog *prog;
 182
 183         meta.seq = seq;
 184         prog = bpf_iter_get_info(&meta, in_stop);
 185         if (!prog)
 186                 return 0;
 187
 188         ctx.meta = &meta;
 189         ctx.task = task;
 190         return bpf_iter_run_prog(prog, &ctx);
 191 }
 192
 193 static int task_seq_show(struct seq_file *seq, void *v)
 194 {
 195         return __task_seq_show(seq, v, false);
 196 }
 197
 198 static void task_seq_stop(struct seq_file *seq, void *v)
 199 {
 200         if (!v)
 201                 (void)__task_seq_show(seq, v, true);
 202         else
 203                 put_task_struct((struct task_struct *)v);
 204 }
 205
 206 static int bpf_iter_attach_task(struct bpf_prog *prog,
 207                                 union bpf_iter_link_info *linfo,
 208                                 struct bpf_iter_aux_info *aux)
 209 {
 210         unsigned int flags;
 211         struct pid *pid;
 212         pid_t tgid;
 213
 214         if ((!!linfo->task.tid + !!linfo->task.pid + !!linfo->task.pid_fd) > 1)
 215                 return -EINVAL;
 216
 217         aux->task.type = BPF_TASK_ITER_ALL;
 218         if (linfo->task.tid != 0) {
 219                 aux->task.type = BPF_TASK_ITER_TID;
 220                 aux->task.pid = linfo->task.tid;
 221         }
 222         if (linfo->task.pid != 0) {
 223                 aux->task.type = BPF_TASK_ITER_TGID;
 224                 aux->task.pid = linfo->task.pid;
 225         }
 226         if (linfo->task.pid_fd != 0) {
 227                 aux->task.type = BPF_TASK_ITER_TGID;
 228
 229                 pid = pidfd_get_pid(linfo->task.pid_fd, &flags);
 230                 if (IS_ERR(pid))
 231                         return PTR_ERR(pid);
 232
 233                 tgid = pid_nr_ns(pid, task_active_pid_ns(current));
 234                 aux->task.pid = tgid;
 235                 put_pid(pid);
 236         }
 237
 238         return 0;
 239 }
 240
 241 static const struct seq_operations task_seq_ops = {
 242         .start  = task_seq_start,
 243         .next   = task_seq_next,
 244         .stop   = task_seq_stop,
 245         .show   = task_seq_show,
 246 };
 247
 248 struct bpf_iter_seq_task_file_info {
 249         /* The first field must be struct bpf_iter_seq_task_common.
 250          * this is assumed by {init, fini}_seq_pidns() callback functions.
 251          */
 252         struct bpf_iter_seq_task_common common;
 253         struct task_struct *task;
 254         u32 tid;
 255         u32 fd;
 256 };
 257
 258 static struct file *
 259 task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info)
 260 {
 261         u32 saved_tid = info->tid;
 262         struct task_struct *curr_task;
 263         unsigned int curr_fd = info->fd;
 264         struct file *f;
 265
 266         /* If this function returns a non-NULL file object,
 267          * it held a reference to the task/file.
 268          * Otherwise, it does not hold any reference.
 269          */
 270 again:
 271         if (info->task) {
 272                 curr_task = info->task;
 273                 curr_fd = info->fd;
 274         } else {
 275                 curr_task = task_seq_get_next(&info->common, &info->tid, true);
 276                 if (!curr_task) {
 277                         info->task = NULL;
 278                         return NULL;
 279                 }
 280
 281                 /* set info->task */
 282                 info->task = curr_task;
 283                 if (saved_tid == info->tid)
 284                         curr_fd = info->fd;
 285                 else
 286                         curr_fd = 0;
 287         }
 288
 289         rcu_read_lock();
 290         f = task_lookup_next_fdget_rcu(curr_task, &curr_fd);
 291         if (f) {
 292                 /* set info->fd */
 293                 info->fd = curr_fd;
 294                 rcu_read_unlock();
 295                 return f;
 296         }
 297
 298         /* the current task is done, go to the next task */
 299         rcu_read_unlock();
 300         put_task_struct(curr_task);
 301
 302         if (info->common.type == BPF_TASK_ITER_TID) {
 303                 info->task = NULL;
 304                 return NULL;
 305         }
 306
 307         info->task = NULL;
 308         info->fd = 0;
 309         saved_tid = ++(info->tid);
 310         goto again;
 311 }
 312
 313 static void *task_file_seq_start(struct seq_file *seq, loff_t *pos)
 314 {
 315         struct bpf_iter_seq_task_file_info *info = seq->private;
 316         struct file *file;
 317
 318         info->task = NULL;
 319         file = task_file_seq_get_next(info);
 320         if (file && *pos == 0)
 321                 ++*pos;
 322
 323         return file;
 324 }
 325
 326 static void *task_file_seq_next(struct seq_file *seq, void *v, loff_t *pos)
 327 {
 328         struct bpf_iter_seq_task_file_info *info = seq->private;
 329
 330         ++*pos;
 331         ++info->fd;
 332         fput((struct file *)v);
 333         return task_file_seq_get_next(info);
 334 }
 335
 336 struct bpf_iter__task_file {
 337         __bpf_md_ptr(struct bpf_iter_meta *, meta);
 338         __bpf_md_ptr(struct task_struct *, task);
 339         u32 fd __aligned(8);
 340         __bpf_md_ptr(struct file *, file);
 341 };
 342
 343 DEFINE_BPF_ITER_FUNC(task_file, struct bpf_iter_meta *meta,
 344                      struct task_struct *task, u32 fd,
 345                      struct file *file)
 346
 347 static int __task_file_seq_show(struct seq_file *seq, struct file *file,
 348                                 bool in_stop)
 349 {
 350         struct bpf_iter_seq_task_file_info *info = seq->private;
 351         struct bpf_iter__task_file ctx;
 352         struct bpf_iter_meta meta;
 353         struct bpf_prog *prog;
 354
 355         meta.seq = seq;
 356         prog = bpf_iter_get_info(&meta, in_stop);
 357         if (!prog)
 358                 return 0;
 359
 360         ctx.meta = &meta;
 361         ctx.task = info->task;
 362         ctx.fd = info->fd;
 363         ctx.file = file;
 364         return bpf_iter_run_prog(prog, &ctx);
 365 }
 366
 367 static int task_file_seq_show(struct seq_file *seq, void *v)
 368 {
 369         return __task_file_seq_show(seq, v, false);
 370 }
 371
 372 static void task_file_seq_stop(struct seq_file *seq, void *v)
 373 {
 374         struct bpf_iter_seq_task_file_info *info = seq->private;
 375
 376         if (!v) {
 377                 (void)__task_file_seq_show(seq, v, true);
 378         } else {
 379                 fput((struct file *)v);
 380                 put_task_struct(info->task);
 381                 info->task = NULL;
 382         }
 383 }
 384
 385 static int init_seq_pidns(void *priv_data, struct bpf_iter_aux_info *aux)
 386 {
 387         struct bpf_iter_seq_task_common *common = priv_data;
 388
 389         common->ns = get_pid_ns(task_active_pid_ns(current));
 390         common->type = aux->task.type;
 391         common->pid = aux->task.pid;
 392
 393         return 0;
 394 }
 395
 396 static void fini_seq_pidns(void *priv_data)
 397 {
 398         struct bpf_iter_seq_task_common *common = priv_data;
 399
 400         put_pid_ns(common->ns);
 401 }
 402
 403 static const struct seq_operations task_file_seq_ops = {
 404         .start  = task_file_seq_start,
 405         .next   = task_file_seq_next,
 406         .stop   = task_file_seq_stop,
 407         .show   = task_file_seq_show,
 408 };
 409
 410 struct bpf_iter_seq_task_vma_info {
 411         /* The first field must be struct bpf_iter_seq_task_common.
 412          * this is assumed by {init, fini}_seq_pidns() callback functions.
 413          */
 414         struct bpf_iter_seq_task_common common;
 415         struct task_struct *task;
 416         struct mm_struct *mm;
 417         struct vm_area_struct *vma;
 418         u32 tid;
 419         unsigned long prev_vm_start;
 420         unsigned long prev_vm_end;
 421 };
 422
 423 enum bpf_task_vma_iter_find_op {
 424         task_vma_iter_first_vma,   /* use find_vma() with addr 0 */
 425         task_vma_iter_next_vma,    /* use vma_next() with curr_vma */
 426         task_vma_iter_find_vma,    /* use find_vma() to find next vma */
 427 };
 428
 429 static struct vm_area_struct *
 430 task_vma_seq_get_next(struct bpf_iter_seq_task_vma_info *info)
 431 {
 432         enum bpf_task_vma_iter_find_op op;
 433         struct vm_area_struct *curr_vma;
 434         struct task_struct *curr_task;
 435         struct mm_struct *curr_mm;
 436         u32 saved_tid = info->tid;
 437
 438         /* If this function returns a non-NULL vma, it holds a reference to
 439          * the task_struct, holds a refcount on mm->mm_users, and holds
 440          * read lock on vma->mm->mmap_lock.
 441          * If this function returns NULL, it does not hold any reference or
 442          * lock.
 443          */
 444         if (info->task) {
 445                 curr_task = info->task;
 446                 curr_vma = info->vma;
 447                 curr_mm = info->mm;
 448                 /* In case of lock contention, drop mmap_lock to unblock
 449                  * the writer.
 450                  *
 451                  * After relock, call find(mm, prev_vm_end - 1) to find
 452                  * new vma to process.
 453                  *
 454                  *   +------+------+-----------+
 455                  *   | VMA1 | VMA2 | VMA3      |
 456                  *   +------+------+-----------+
 457                  *   |      |      |           |
 458                  *  4k     8k     16k         400k
 459                  *
 460                  * For example, curr_vma == VMA2. Before unlock, we set
 461                  *
 462                  *    prev_vm_start = 8k
 463                  *    prev_vm_end   = 16k
 464                  *
 465                  * There are a few cases:
 466                  *
 467                  * 1) VMA2 is freed, but VMA3 exists.
 468                  *
 469                  *    find_vma() will return VMA3, just process VMA3.
 470                  *
 471                  * 2) VMA2 still exists.
 472                  *
 473                  *    find_vma() will return VMA2, process VMA2->next.
 474                  *
 475                  * 3) no more vma in this mm.
 476                  *
 477                  *    Process the next task.
 478                  *
 479                  * 4) find_vma() returns a different vma, VMA2'.
 480                  *
 481                  *    4.1) If VMA2 covers same range as VMA2', skip VMA2',
 482                  *         because we already covered the range;
 483                  *    4.2) VMA2 and VMA2' covers different ranges, process
 484                  *         VMA2'.
 485                  */
 486                 if (mmap_lock_is_contended(curr_mm)) {
 487                         info->prev_vm_start = curr_vma->vm_start;
 488                         info->prev_vm_end = curr_vma->vm_end;
 489                         op = task_vma_iter_find_vma;
 490                         mmap_read_unlock(curr_mm);
 491                         if (mmap_read_lock_killable(curr_mm)) {
 492                                 mmput(curr_mm);
 493                                 goto finish;
 494                         }
 495                 } else {
 496                         op = task_vma_iter_next_vma;
 497                 }
 498         } else {
 499 again:
 500                 curr_task = task_seq_get_next(&info->common, &info->tid, true);
 501                 if (!curr_task) {
 502                         info->tid++;
 503                         goto finish;
 504                 }
 505
 506                 if (saved_tid != info->tid) {
 507                         /* new task, process the first vma */
 508                         op = task_vma_iter_first_vma;
 509                 } else {
 510                         /* Found the same tid, which means the user space
 511                          * finished data in previous buffer and read more.
 512                          * We dropped mmap_lock before returning to user
 513                          * space, so it is necessary to use find_vma() to
 514                          * find the next vma to process.
 515                          */
 516                         op = task_vma_iter_find_vma;
 517                 }
 518
 519                 curr_mm = get_task_mm(curr_task);
 520                 if (!curr_mm)
 521                         goto next_task;
 522
 523                 if (mmap_read_lock_killable(curr_mm)) {
 524                         mmput(curr_mm);
 525                         goto finish;
 526                 }
 527         }
 528
 529         switch (op) {
 530         case task_vma_iter_first_vma:
 531                 curr_vma = find_vma(curr_mm, 0);
 532                 break;
 533         case task_vma_iter_next_vma:
 534                 curr_vma = find_vma(curr_mm, curr_vma->vm_end);
 535                 break;
 536         case task_vma_iter_find_vma:
 537                 /* We dropped mmap_lock so it is necessary to use find_vma
 538                  * to find the next vma. This is similar to the  mechanism
 539                  * in show_smaps_rollup().
 540                  */
 541                 curr_vma = find_vma(curr_mm, info->prev_vm_end - 1);
 542                 /* case 1) and 4.2) above just use curr_vma */
 543
 544                 /* check for case 2) or case 4.1) above */
 545                 if (curr_vma &&
 546                     curr_vma->vm_start == info->prev_vm_start &&
 547                     curr_vma->vm_end == info->prev_vm_end)
 548                         curr_vma = find_vma(curr_mm, curr_vma->vm_end);
 549                 break;
 550         }
 551         if (!curr_vma) {
 552                 /* case 3) above, or case 2) 4.1) with vma->next == NULL */
 553                 mmap_read_unlock(curr_mm);
 554                 mmput(curr_mm);
 555                 goto next_task;
 556         }
 557         info->task = curr_task;
 558         info->vma = curr_vma;
 559         info->mm = curr_mm;
 560         return curr_vma;
 561
 562 next_task:
 563         if (info->common.type == BPF_TASK_ITER_TID)
 564                 goto finish;
 565
 566         put_task_struct(curr_task);
 567         info->task = NULL;
 568         info->mm = NULL;
 569         info->tid++;
 570         goto again;
 571
 572 finish:
 573         if (curr_task)
 574                 put_task_struct(curr_task);
 575         info->task = NULL;
 576         info->vma = NULL;
 577         info->mm = NULL;
 578         return NULL;
 579 }
 580
 581 static void *task_vma_seq_start(struct seq_file *seq, loff_t *pos)
 582 {
 583         struct bpf_iter_seq_task_vma_info *info = seq->private;
 584         struct vm_area_struct *vma;
 585
 586         vma = task_vma_seq_get_next(info);
 587         if (vma && *pos == 0)
 588                 ++*pos;
 589
 590         return vma;
 591 }
 592
 593 static void *task_vma_seq_next(struct seq_file *seq, void *v, loff_t *pos)
 594 {
 595         struct bpf_iter_seq_task_vma_info *info = seq->private;
 596
 597         ++*pos;
 598         return task_vma_seq_get_next(info);
 599 }
 600
 601 struct bpf_iter__task_vma {
 602         __bpf_md_ptr(struct bpf_iter_meta *, meta);
 603         __bpf_md_ptr(struct task_struct *, task);
 604         __bpf_md_ptr(struct vm_area_struct *, vma);
 605 };
 606
 607 DEFINE_BPF_ITER_FUNC(task_vma, struct bpf_iter_meta *meta,
 608                      struct task_struct *task, struct vm_area_struct *vma)
 609
 610 static int __task_vma_seq_show(struct seq_file *seq, bool in_stop)
 611 {
 612         struct bpf_iter_seq_task_vma_info *info = seq->private;
 613         struct bpf_iter__task_vma ctx;
 614         struct bpf_iter_meta meta;
 615         struct bpf_prog *prog;
 616
 617         meta.seq = seq;
 618         prog = bpf_iter_get_info(&meta, in_stop);
 619         if (!prog)
 620                 return 0;
 621
 622         ctx.meta = &meta;
 623         ctx.task = info->task;
 624         ctx.vma = info->vma;
 625         return bpf_iter_run_prog(prog, &ctx);
 626 }
 627
 628 static int task_vma_seq_show(struct seq_file *seq, void *v)
 629 {
 630         return __task_vma_seq_show(seq, false);
 631 }
 632
 633 static void task_vma_seq_stop(struct seq_file *seq, void *v)
 634 {
 635         struct bpf_iter_seq_task_vma_info *info = seq->private;
 636
 637         if (!v) {
 638                 (void)__task_vma_seq_show(seq, true);
 639         } else {
 640                 /* info->vma has not been seen by the BPF program. If the
 641                  * user space reads more, task_vma_seq_get_next should
 642                  * return this vma again. Set prev_vm_start to ~0UL,
 643                  * so that we don't skip the vma returned by the next
 644                  * find_vma() (case task_vma_iter_find_vma in
 645                  * task_vma_seq_get_next()).
 646                  */
 647                 info->prev_vm_start = ~0UL;
 648                 info->prev_vm_end = info->vma->vm_end;
 649                 mmap_read_unlock(info->mm);
 650                 mmput(info->mm);
 651                 info->mm = NULL;
 652                 put_task_struct(info->task);
 653                 info->task = NULL;
 654         }
 655 }
 656
 657 static const struct seq_operations task_vma_seq_ops = {
 658         .start  = task_vma_seq_start,
 659         .next   = task_vma_seq_next,
 660         .stop   = task_vma_seq_stop,
 661         .show   = task_vma_seq_show,
 662 };
 663
 664 static const struct bpf_iter_seq_info task_seq_info = {
 665         .seq_ops                = &task_seq_ops,
 666         .init_seq_private       = init_seq_pidns,
 667         .fini_seq_private       = fini_seq_pidns,
 668         .seq_priv_size          = sizeof(struct bpf_iter_seq_task_info),
 669 };
 670
 671 static int bpf_iter_fill_link_info(const struct bpf_iter_aux_info *aux, struct bpf_link_info *info)
 672 {
 673         switch (aux->task.type) {
 674         case BPF_TASK_ITER_TID:
 675                 info->iter.task.tid = aux->task.pid;
 676                 break;
 677         case BPF_TASK_ITER_TGID:
 678                 info->iter.task.pid = aux->task.pid;
 679                 break;
 680         default:
 681                 break;
 682         }
 683         return 0;
 684 }
 685
 686 static void bpf_iter_task_show_fdinfo(const struct bpf_iter_aux_info *aux, struct seq_file *seq)
 687 {
 688         seq_printf(seq, "task_type:\t%s\n", iter_task_type_names[aux->task.type]);
 689         if (aux->task.type == BPF_TASK_ITER_TID)
 690                 seq_printf(seq, "tid:\t%u\n", aux->task.pid);
 691         else if (aux->task.type == BPF_TASK_ITER_TGID)
 692                 seq_printf(seq, "pid:\t%u\n", aux->task.pid);
 693 }
 694
 695 static struct bpf_iter_reg task_reg_info = {
 696         .target                 = "task",
 697         .attach_target          = bpf_iter_attach_task,
 698         .feature                = BPF_ITER_RESCHED,
 699         .ctx_arg_info_size      = 1,
 700         .ctx_arg_info           = {
 701                 { offsetof(struct bpf_iter__task, task),
 702                   PTR_TO_BTF_ID_OR_NULL | PTR_TRUSTED },
 703         },
 704         .seq_info               = &task_seq_info,
 705         .fill_link_info         = bpf_iter_fill_link_info,
 706         .show_fdinfo            = bpf_iter_task_show_fdinfo,
 707 };
 708
 709 static const struct bpf_iter_seq_info task_file_seq_info = {
 710         .seq_ops                = &task_file_seq_ops,
 711         .init_seq_private       = init_seq_pidns,
 712         .fini_seq_private       = fini_seq_pidns,
 713         .seq_priv_size          = sizeof(struct bpf_iter_seq_task_file_info),
 714 };
 715
 716 static struct bpf_iter_reg task_file_reg_info = {
 717         .target                 = "task_file",
 718         .attach_target          = bpf_iter_attach_task,
 719         .feature                = BPF_ITER_RESCHED,
 720         .ctx_arg_info_size      = 2,
 721         .ctx_arg_info           = {
 722                 { offsetof(struct bpf_iter__task_file, task),
 723                   PTR_TO_BTF_ID_OR_NULL },
 724                 { offsetof(struct bpf_iter__task_file, file),
 725                   PTR_TO_BTF_ID_OR_NULL },
 726         },
 727         .seq_info               = &task_file_seq_info,
 728         .fill_link_info         = bpf_iter_fill_link_info,
 729         .show_fdinfo            = bpf_iter_task_show_fdinfo,
 730 };
 731
 732 static const struct bpf_iter_seq_info task_vma_seq_info = {
 733         .seq_ops                = &task_vma_seq_ops,
 734         .init_seq_private       = init_seq_pidns,
 735         .fini_seq_private       = fini_seq_pidns,
 736         .seq_priv_size          = sizeof(struct bpf_iter_seq_task_vma_info),
 737 };
 738
 739 static struct bpf_iter_reg task_vma_reg_info = {
 740         .target                 = "task_vma",
 741         .attach_target          = bpf_iter_attach_task,
 742         .feature                = BPF_ITER_RESCHED,
 743         .ctx_arg_info_size      = 2,
 744         .ctx_arg_info           = {
 745                 { offsetof(struct bpf_iter__task_vma, task),
 746                   PTR_TO_BTF_ID_OR_NULL },
 747                 { offsetof(struct bpf_iter__task_vma, vma),
 748                   PTR_TO_BTF_ID_OR_NULL },
 749         },
 750         .seq_info               = &task_vma_seq_info,
 751         .fill_link_info         = bpf_iter_fill_link_info,
 752         .show_fdinfo            = bpf_iter_task_show_fdinfo,
 753 };
 754
 755 BPF_CALL_5(bpf_find_vma, struct task_struct *, task, u64, start,
 756            bpf_callback_t, callback_fn, void *, callback_ctx, u64, flags)
 757 {
 758         struct mmap_unlock_irq_work *work = NULL;
 759         struct vm_area_struct *vma;
 760         bool irq_work_busy = false;
 761         struct mm_struct *mm;
 762         int ret = -ENOENT;
 763
 764         if (flags)
 765                 return -EINVAL;
 766
 767         if (!task)
 768                 return -ENOENT;
 769
 770         mm = task->mm;
 771         if (!mm)
 772                 return -ENOENT;
 773
 774         irq_work_busy = bpf_mmap_unlock_get_irq_work(&work);
 775
 776         if (irq_work_busy || !mmap_read_trylock(mm))
 777                 return -EBUSY;
 778
 779         vma = find_vma(mm, start);
 780
 781         if (vma && vma->vm_start <= start && vma->vm_end > start) {
 782                 callback_fn((u64)(long)task, (u64)(long)vma,
 783                             (u64)(long)callback_ctx, 0, 0);
 784                 ret = 0;
 785         }
 786         bpf_mmap_unlock_mm(work, mm);
 787         return ret;
 788 }
 789
 790 const struct bpf_func_proto bpf_find_vma_proto = {
 791         .func           = bpf_find_vma,
 792         .ret_type       = RET_INTEGER,
 793         .arg1_type      = ARG_PTR_TO_BTF_ID,
 794         .arg1_btf_id    = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
 795         .arg2_type      = ARG_ANYTHING,
 796         .arg3_type      = ARG_PTR_TO_FUNC,
 797         .arg4_type      = ARG_PTR_TO_STACK_OR_NULL,
 798         .arg5_type      = ARG_ANYTHING,
 799 };
 800
 801 struct bpf_iter_task_vma_kern_data {
 802         struct task_struct *task;
 803         struct mm_struct *mm;
 804         struct mmap_unlock_irq_work *work;
 805         struct vma_iterator vmi;
 806 };
 807
 808 struct bpf_iter_task_vma {
 809         /* opaque iterator state; having __u64 here allows to preserve correct
 810          * alignment requirements in vmlinux.h, generated from BTF
 811          */
 812         __u64 __opaque[1];
 813 } __attribute__((aligned(8)));
 814
 815 /* Non-opaque version of bpf_iter_task_vma */
 816 struct bpf_iter_task_vma_kern {
 817         struct bpf_iter_task_vma_kern_data *data;
 818 } __attribute__((aligned(8)));
 819
 820 __bpf_kfunc_start_defs();
 821
 822 __bpf_kfunc int bpf_iter_task_vma_new(struct bpf_iter_task_vma *it,
 823                                       struct task_struct *task, u64 addr)
 824 {
 825         struct bpf_iter_task_vma_kern *kit = (void *)it;
 826         bool irq_work_busy = false;
 827         int err;
 828
 829         BUILD_BUG_ON(sizeof(struct bpf_iter_task_vma_kern) != sizeof(struct bpf_iter_task_vma));
 830         BUILD_BUG_ON(__alignof__(struct bpf_iter_task_vma_kern) != __alignof__(struct bpf_iter_task_vma));
 831
 832         /* is_iter_reg_valid_uninit guarantees that kit hasn't been initialized
 833          * before, so non-NULL kit->data doesn't point to previously
 834          * bpf_mem_alloc'd bpf_iter_task_vma_kern_data
 835          */
 836         kit->data = bpf_mem_alloc(&bpf_global_ma, sizeof(struct bpf_iter_task_vma_kern_data));
 837         if (!kit->data)
 838                 return -ENOMEM;
 839
 840         kit->data->task = get_task_struct(task);
 841         kit->data->mm = task->mm;
 842         if (!kit->data->mm) {
 843                 err = -ENOENT;
 844                 goto err_cleanup_iter;
 845         }
 846
 847         /* kit->data->work == NULL is valid after bpf_mmap_unlock_get_irq_work */
 848         irq_work_busy = bpf_mmap_unlock_get_irq_work(&kit->data->work);
 849         if (irq_work_busy || !mmap_read_trylock(kit->data->mm)) {
 850                 err = -EBUSY;
 851                 goto err_cleanup_iter;
 852         }
 853
 854         vma_iter_init(&kit->data->vmi, kit->data->mm, addr);
 855         return 0;
 856
 857 err_cleanup_iter:
 858         if (kit->data->task)
 859                 put_task_struct(kit->data->task);
 860         bpf_mem_free(&bpf_global_ma, kit->data);
 861         /* NULL kit->data signals failed bpf_iter_task_vma initialization */
 862         kit->data = NULL;
 863         return err;
 864 }
 865
 866 __bpf_kfunc struct vm_area_struct *bpf_iter_task_vma_next(struct bpf_iter_task_vma *it)
 867 {
 868         struct bpf_iter_task_vma_kern *kit = (void *)it;
 869
 870         if (!kit->data) /* bpf_iter_task_vma_new failed */
 871                 return NULL;
 872         return vma_next(&kit->data->vmi);
 873 }
 874
 875 __bpf_kfunc void bpf_iter_task_vma_destroy(struct bpf_iter_task_vma *it)
 876 {
 877         struct bpf_iter_task_vma_kern *kit = (void *)it;
 878
 879         if (kit->data) {
 880                 bpf_mmap_unlock_mm(kit->data->work, kit->data->mm);
 881                 put_task_struct(kit->data->task);
 882                 bpf_mem_free(&bpf_global_ma, kit->data);
 883         }
 884 }
 885
 886 __bpf_kfunc_end_defs();
 887
 888 #ifdef CONFIG_CGROUPS
 889
 890 struct bpf_iter_css_task {
 891         __u64 __opaque[1];
 892 } __attribute__((aligned(8)));
 893
 894 struct bpf_iter_css_task_kern {
 895         struct css_task_iter *css_it;
 896 } __attribute__((aligned(8)));
 897
 898 __bpf_kfunc_start_defs();
 899
 900 __bpf_kfunc int bpf_iter_css_task_new(struct bpf_iter_css_task *it,
 901                 struct cgroup_subsys_state *css, unsigned int flags)
 902 {
 903         struct bpf_iter_css_task_kern *kit = (void *)it;
 904
 905         BUILD_BUG_ON(sizeof(struct bpf_iter_css_task_kern) != sizeof(struct bpf_iter_css_task));
 906         BUILD_BUG_ON(__alignof__(struct bpf_iter_css_task_kern) !=
 907                                         __alignof__(struct bpf_iter_css_task));
 908         kit->css_it = NULL;
 909         switch (flags) {
 910         case CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED:
 911         case CSS_TASK_ITER_PROCS:
 912         case 0:
 913                 break;
 914         default:
 915                 return -EINVAL;
 916         }
 917
 918         kit->css_it = bpf_mem_alloc(&bpf_global_ma, sizeof(struct css_task_iter));
 919         if (!kit->css_it)
 920                 return -ENOMEM;
 921         css_task_iter_start(css, flags, kit->css_it);
 922         return 0;
 923 }
 924
 925 __bpf_kfunc struct task_struct *bpf_iter_css_task_next(struct bpf_iter_css_task *it)
 926 {
 927         struct bpf_iter_css_task_kern *kit = (void *)it;
 928
 929         if (!kit->css_it)
 930                 return NULL;
 931         return css_task_iter_next(kit->css_it);
 932 }
 933
 934 __bpf_kfunc void bpf_iter_css_task_destroy(struct bpf_iter_css_task *it)
 935 {
 936         struct bpf_iter_css_task_kern *kit = (void *)it;
 937
 938         if (!kit->css_it)
 939                 return;
 940         css_task_iter_end(kit->css_it);
 941         bpf_mem_free(&bpf_global_ma, kit->css_it);
 942 }
 943
 944 __bpf_kfunc_end_defs();
 945
 946 #endif /* CONFIG_CGROUPS */
 947
 948 struct bpf_iter_task {
 949         __u64 __opaque[3];
 950 } __attribute__((aligned(8)));
 951
 952 struct bpf_iter_task_kern {
 953         struct task_struct *task;
 954         struct task_struct *pos;
 955         unsigned int flags;
 956 } __attribute__((aligned(8)));
 957
 958 enum {
 959         /* all process in the system */
 960         BPF_TASK_ITER_ALL_PROCS,
 961         /* all threads in the system */
 962         BPF_TASK_ITER_ALL_THREADS,
 963         /* all threads of a specific process */
 964         BPF_TASK_ITER_PROC_THREADS
 965 };
 966
 967 __bpf_kfunc_start_defs();
 968
 969 __bpf_kfunc int bpf_iter_task_new(struct bpf_iter_task *it,
 970                 struct task_struct *task__nullable, unsigned int flags)
 971 {
 972         struct bpf_iter_task_kern *kit = (void *)it;
 973
 974         BUILD_BUG_ON(sizeof(struct bpf_iter_task_kern) > sizeof(struct bpf_iter_task));
 975         BUILD_BUG_ON(__alignof__(struct bpf_iter_task_kern) !=
 976                                         __alignof__(struct bpf_iter_task));
 977
 978         kit->pos = NULL;
 979
 980         switch (flags) {
 981         case BPF_TASK_ITER_ALL_THREADS:
 982         case BPF_TASK_ITER_ALL_PROCS:
 983                 break;
 984         case BPF_TASK_ITER_PROC_THREADS:
 985                 if (!task__nullable)
 986                         return -EINVAL;
 987                 break;
 988         default:
 989                 return -EINVAL;
 990         }
 991
 992         if (flags == BPF_TASK_ITER_PROC_THREADS)
 993                 kit->task = task__nullable;
 994         else
 995                 kit->task = &init_task;
 996         kit->pos = kit->task;
 997         kit->flags = flags;
 998         return 0;
 999 }
1000
1001 __bpf_kfunc struct task_struct *bpf_iter_task_next(struct bpf_iter_task *it)
1002 {
1003         struct bpf_iter_task_kern *kit = (void *)it;
1004         struct task_struct *pos;
1005         unsigned int flags;
1006
1007         flags = kit->flags;
1008         pos = kit->pos;
1009
1010         if (!pos)
1011                 return pos;
1012
1013         if (flags == BPF_TASK_ITER_ALL_PROCS)
1014                 goto get_next_task;
1015
1016         kit->pos = __next_thread(kit->pos);
1017         if (kit->pos || flags == BPF_TASK_ITER_PROC_THREADS)
1018                 return pos;
1019
1020 get_next_task:
1021         kit->task = next_task(kit->task);
1022         if (kit->task == &init_task)
1023                 kit->pos = NULL;
1024         else
1025                 kit->pos = kit->task;
1026
1027         return pos;
1028 }
1029
1030 __bpf_kfunc void bpf_iter_task_destroy(struct bpf_iter_task *it)
1031 {
1032 }
1033
1034 __bpf_kfunc_end_defs();
1035
1036 DEFINE_PER_CPU(struct mmap_unlock_irq_work, mmap_unlock_work);
1037
1038 static void do_mmap_read_unlock(struct irq_work *entry)
1039 {
1040         struct mmap_unlock_irq_work *work;
1041
1042         if (WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_RT)))
1043                 return;
1044
1045         work = container_of(entry, struct mmap_unlock_irq_work, irq_work);
1046         mmap_read_unlock_non_owner(work->mm);
1047 }
1048
1049 static int __init task_iter_init(void)
1050 {
1051         struct mmap_unlock_irq_work *work;
1052         int ret, cpu;
1053
1054         for_each_possible_cpu(cpu) {
1055                 work = per_cpu_ptr(&mmap_unlock_work, cpu);
1056                 init_irq_work(&work->irq_work, do_mmap_read_unlock);
1057         }
1058
1059         task_reg_info.ctx_arg_info[0].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_TASK];
1060         ret = bpf_iter_reg_target(&task_reg_info);
1061         if (ret)
1062                 return ret;
1063
1064         task_file_reg_info.ctx_arg_info[0].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_TASK];
1065         task_file_reg_info.ctx_arg_info[1].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_FILE];
1066         ret =  bpf_iter_reg_target(&task_file_reg_info);
1067         if (ret)
1068                 return ret;
1069
1070         task_vma_reg_info.ctx_arg_info[0].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_TASK];
1071         task_vma_reg_info.ctx_arg_info[1].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA];
1072         return bpf_iter_reg_target(&task_vma_reg_info);
1073 }
1074 late_initcall(task_iter_init);