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
  75         next_tid = __task_pid_nr_ns(task, PIDTYPE_PID, common->ns);
  76         if (!next_tid || next_tid == common->pid) {
  77                 /* Run out of tasks of a process.  The tasks of a
  78                  * thread_group are linked as circular linked list.
  79                  */
  80                 return NULL;
  81         }
  82
  83         if (skip_if_dup_files && task->files == task->group_leader->files)
  84                 goto retry;
  85
  86         *tid = common->pid_visiting = next_tid;
  87         get_task_struct(task);
  88         return task;
  89 }
  90
  91 static struct task_struct *task_seq_get_next(struct bpf_iter_seq_task_common *common,
  92                                              u32 *tid,
  93                                              bool skip_if_dup_files)
  94 {
  95         struct task_struct *task = NULL;
  96         struct pid *pid;
  97
  98         if (common->type == BPF_TASK_ITER_TID) {
  99                 if (*tid && *tid != common->pid)
 100                         return NULL;
 101                 rcu_read_lock();
 102                 pid = find_pid_ns(common->pid, common->ns);
 103                 if (pid) {
 104                         task = get_pid_task(pid, PIDTYPE_TGID);
 105                         *tid = common->pid;
 106                 }
 107                 rcu_read_unlock();
 108
 109                 return task;
 110         }
 111
 112         if (common->type == BPF_TASK_ITER_TGID) {
 113                 rcu_read_lock();
 114                 task = task_group_seq_get_next(common, tid, skip_if_dup_files);
 115                 rcu_read_unlock();
 116
 117                 return task;
 118         }
 119
 120         rcu_read_lock();
 121 retry:
 122         pid = find_ge_pid(*tid, common->ns);
 123         if (pid) {
 124                 *tid = pid_nr_ns(pid, common->ns);
 125                 task = get_pid_task(pid, PIDTYPE_PID);
 126                 if (!task) {
 127                         ++*tid;
 128                         goto retry;
 129                 } else if (skip_if_dup_files && !thread_group_leader(task) &&
 130                            task->files == task->group_leader->files) {
 131                         put_task_struct(task);
 132                         task = NULL;
 133                         ++*tid;
 134                         goto retry;
 135                 }
 136         }
 137         rcu_read_unlock();
 138
 139         return task;
 140 }
 141
 142 static void *task_seq_start(struct seq_file *seq, loff_t *pos)
 143 {
 144         struct bpf_iter_seq_task_info *info = seq->private;
 145         struct task_struct *task;
 146
 147         task = task_seq_get_next(&info->common, &info->tid, false);
 148         if (!task)
 149                 return NULL;
 150
 151         if (*pos == 0)
 152                 ++*pos;
 153         return task;
 154 }
 155
 156 static void *task_seq_next(struct seq_file *seq, void *v, loff_t *pos)
 157 {
 158         struct bpf_iter_seq_task_info *info = seq->private;
 159         struct task_struct *task;
 160
 161         ++*pos;
 162         ++info->tid;
 163         put_task_struct((struct task_struct *)v);
 164         task = task_seq_get_next(&info->common, &info->tid, false);
 165         if (!task)
 166                 return NULL;
 167
 168         return task;
 169 }
 170
 171 struct bpf_iter__task {
 172         __bpf_md_ptr(struct bpf_iter_meta *, meta);
 173         __bpf_md_ptr(struct task_struct *, task);
 174 };
 175
 176 DEFINE_BPF_ITER_FUNC(task, struct bpf_iter_meta *meta, struct task_struct *task)
 177
 178 static int __task_seq_show(struct seq_file *seq, struct task_struct *task,
 179                            bool in_stop)
 180 {
 181         struct bpf_iter_meta meta;
 182         struct bpf_iter__task ctx;
 183         struct bpf_prog *prog;
 184
 185         meta.seq = seq;
 186         prog = bpf_iter_get_info(&meta, in_stop);
 187         if (!prog)
 188                 return 0;
 189
 190         ctx.meta = &meta;
 191         ctx.task = task;
 192         return bpf_iter_run_prog(prog, &ctx);
 193 }
 194
 195 static int task_seq_show(struct seq_file *seq, void *v)
 196 {
 197         return __task_seq_show(seq, v, false);
 198 }
 199
 200 static void task_seq_stop(struct seq_file *seq, void *v)
 201 {
 202         if (!v)
 203                 (void)__task_seq_show(seq, v, true);
 204         else
 205                 put_task_struct((struct task_struct *)v);
 206 }
 207
 208 static int bpf_iter_attach_task(struct bpf_prog *prog,
 209                                 union bpf_iter_link_info *linfo,
 210                                 struct bpf_iter_aux_info *aux)
 211 {
 212         unsigned int flags;
 213         struct pid *pid;
 214         pid_t tgid;
 215
 216         if ((!!linfo->task.tid + !!linfo->task.pid + !!linfo->task.pid_fd) > 1)
 217                 return -EINVAL;
 218
 219         aux->task.type = BPF_TASK_ITER_ALL;
 220         if (linfo->task.tid != 0) {
 221                 aux->task.type = BPF_TASK_ITER_TID;
 222                 aux->task.pid = linfo->task.tid;
 223         }
 224         if (linfo->task.pid != 0) {
 225                 aux->task.type = BPF_TASK_ITER_TGID;
 226                 aux->task.pid = linfo->task.pid;
 227         }
 228         if (linfo->task.pid_fd != 0) {
 229                 aux->task.type = BPF_TASK_ITER_TGID;
 230
 231                 pid = pidfd_get_pid(linfo->task.pid_fd, &flags);
 232                 if (IS_ERR(pid))
 233                         return PTR_ERR(pid);
 234
 235                 tgid = pid_nr_ns(pid, task_active_pid_ns(current));
 236                 aux->task.pid = tgid;
 237                 put_pid(pid);
 238         }
 239
 240         return 0;
 241 }
 242
 243 static const struct seq_operations task_seq_ops = {
 244         .start  = task_seq_start,
 245         .next   = task_seq_next,
 246         .stop   = task_seq_stop,
 247         .show   = task_seq_show,
 248 };
 249
 250 struct bpf_iter_seq_task_file_info {
 251         /* The first field must be struct bpf_iter_seq_task_common.
 252          * this is assumed by {init, fini}_seq_pidns() callback functions.
 253          */
 254         struct bpf_iter_seq_task_common common;
 255         struct task_struct *task;
 256         u32 tid;
 257         u32 fd;
 258 };
 259
 260 static struct file *
 261 task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info)
 262 {
 263         u32 saved_tid = info->tid;
 264         struct task_struct *curr_task;
 265         unsigned int curr_fd = info->fd;
 266
 267         /* If this function returns a non-NULL file object,
 268          * it held a reference to the task/file.
 269          * Otherwise, it does not hold any reference.
 270          */
 271 again:
 272         if (info->task) {
 273                 curr_task = info->task;
 274                 curr_fd = info->fd;
 275         } else {
 276                 curr_task = task_seq_get_next(&info->common, &info->tid, true);
 277                 if (!curr_task) {
 278                         info->task = NULL;
 279                         return NULL;
 280                 }
 281
 282                 /* set info->task */
 283                 info->task = curr_task;
 284                 if (saved_tid == info->tid)
 285                         curr_fd = info->fd;
 286                 else
 287                         curr_fd = 0;
 288         }
 289
 290         rcu_read_lock();
 291         for (;; curr_fd++) {
 292                 struct file *f;
 293                 f = task_lookup_next_fdget_rcu(curr_task, &curr_fd);
 294                 if (!f)
 295                         break;
 296
 297                 /* set info->fd */
 298                 info->fd = curr_fd;
 299                 rcu_read_unlock();
 300                 return f;
 301         }
 302
 303         /* the current task is done, go to the next task */
 304         rcu_read_unlock();
 305         put_task_struct(curr_task);
 306
 307         if (info->common.type == BPF_TASK_ITER_TID) {
 308                 info->task = NULL;
 309                 return NULL;
 310         }
 311
 312         info->task = NULL;
 313         info->fd = 0;
 314         saved_tid = ++(info->tid);
 315         goto again;
 316 }
 317
 318 static void *task_file_seq_start(struct seq_file *seq, loff_t *pos)
 319 {
 320         struct bpf_iter_seq_task_file_info *info = seq->private;
 321         struct file *file;
 322
 323         info->task = NULL;
 324         file = task_file_seq_get_next(info);
 325         if (file && *pos == 0)
 326                 ++*pos;
 327
 328         return file;
 329 }
 330
 331 static void *task_file_seq_next(struct seq_file *seq, void *v, loff_t *pos)
 332 {
 333         struct bpf_iter_seq_task_file_info *info = seq->private;
 334
 335         ++*pos;
 336         ++info->fd;
 337         fput((struct file *)v);
 338         return task_file_seq_get_next(info);
 339 }
 340
 341 struct bpf_iter__task_file {
 342         __bpf_md_ptr(struct bpf_iter_meta *, meta);
 343         __bpf_md_ptr(struct task_struct *, task);
 344         u32 fd __aligned(8);
 345         __bpf_md_ptr(struct file *, file);
 346 };
 347
 348 DEFINE_BPF_ITER_FUNC(task_file, struct bpf_iter_meta *meta,
 349                      struct task_struct *task, u32 fd,
 350                      struct file *file)
 351
 352 static int __task_file_seq_show(struct seq_file *seq, struct file *file,
 353                                 bool in_stop)
 354 {
 355         struct bpf_iter_seq_task_file_info *info = seq->private;
 356         struct bpf_iter__task_file ctx;
 357         struct bpf_iter_meta meta;
 358         struct bpf_prog *prog;
 359
 360         meta.seq = seq;
 361         prog = bpf_iter_get_info(&meta, in_stop);
 362         if (!prog)
 363                 return 0;
 364
 365         ctx.meta = &meta;
 366         ctx.task = info->task;
 367         ctx.fd = info->fd;
 368         ctx.file = file;
 369         return bpf_iter_run_prog(prog, &ctx);
 370 }
 371
 372 static int task_file_seq_show(struct seq_file *seq, void *v)
 373 {
 374         return __task_file_seq_show(seq, v, false);
 375 }
 376
 377 static void task_file_seq_stop(struct seq_file *seq, void *v)
 378 {
 379         struct bpf_iter_seq_task_file_info *info = seq->private;
 380
 381         if (!v) {
 382                 (void)__task_file_seq_show(seq, v, true);
 383         } else {
 384                 fput((struct file *)v);
 385                 put_task_struct(info->task);
 386                 info->task = NULL;
 387         }
 388 }
 389
 390 static int init_seq_pidns(void *priv_data, struct bpf_iter_aux_info *aux)
 391 {
 392         struct bpf_iter_seq_task_common *common = priv_data;
 393
 394         common->ns = get_pid_ns(task_active_pid_ns(current));
 395         common->type = aux->task.type;
 396         common->pid = aux->task.pid;
 397
 398         return 0;
 399 }
 400
 401 static void fini_seq_pidns(void *priv_data)
 402 {
 403         struct bpf_iter_seq_task_common *common = priv_data;
 404
 405         put_pid_ns(common->ns);
 406 }
 407
 408 static const struct seq_operations task_file_seq_ops = {
 409         .start  = task_file_seq_start,
 410         .next   = task_file_seq_next,
 411         .stop   = task_file_seq_stop,
 412         .show   = task_file_seq_show,
 413 };
 414
 415 struct bpf_iter_seq_task_vma_info {
 416         /* The first field must be struct bpf_iter_seq_task_common.
 417          * this is assumed by {init, fini}_seq_pidns() callback functions.
 418          */
 419         struct bpf_iter_seq_task_common common;
 420         struct task_struct *task;
 421         struct mm_struct *mm;
 422         struct vm_area_struct *vma;
 423         u32 tid;
 424         unsigned long prev_vm_start;
 425         unsigned long prev_vm_end;
 426 };
 427
 428 enum bpf_task_vma_iter_find_op {
 429         task_vma_iter_first_vma,   /* use find_vma() with addr 0 */
 430         task_vma_iter_next_vma,    /* use vma_next() with curr_vma */
 431         task_vma_iter_find_vma,    /* use find_vma() to find next vma */
 432 };
 433
 434 static struct vm_area_struct *
 435 task_vma_seq_get_next(struct bpf_iter_seq_task_vma_info *info)
 436 {
 437         enum bpf_task_vma_iter_find_op op;
 438         struct vm_area_struct *curr_vma;
 439         struct task_struct *curr_task;
 440         struct mm_struct *curr_mm;
 441         u32 saved_tid = info->tid;
 442
 443         /* If this function returns a non-NULL vma, it holds a reference to
 444          * the task_struct, holds a refcount on mm->mm_users, and holds
 445          * read lock on vma->mm->mmap_lock.
 446          * If this function returns NULL, it does not hold any reference or
 447          * lock.
 448          */
 449         if (info->task) {
 450                 curr_task = info->task;
 451                 curr_vma = info->vma;
 452                 curr_mm = info->mm;
 453                 /* In case of lock contention, drop mmap_lock to unblock
 454                  * the writer.
 455                  *
 456                  * After relock, call find(mm, prev_vm_end - 1) to find
 457                  * new vma to process.
 458                  *
 459                  *   +------+------+-----------+
 460                  *   | VMA1 | VMA2 | VMA3      |
 461                  *   +------+------+-----------+
 462                  *   |      |      |           |
 463                  *  4k     8k     16k         400k
 464                  *
 465                  * For example, curr_vma == VMA2. Before unlock, we set
 466                  *
 467                  *    prev_vm_start = 8k
 468                  *    prev_vm_end   = 16k
 469                  *
 470                  * There are a few cases:
 471                  *
 472                  * 1) VMA2 is freed, but VMA3 exists.
 473                  *
 474                  *    find_vma() will return VMA3, just process VMA3.
 475                  *
 476                  * 2) VMA2 still exists.
 477                  *
 478                  *    find_vma() will return VMA2, process VMA2->next.
 479                  *
 480                  * 3) no more vma in this mm.
 481                  *
 482                  *    Process the next task.
 483                  *
 484                  * 4) find_vma() returns a different vma, VMA2'.
 485                  *
 486                  *    4.1) If VMA2 covers same range as VMA2', skip VMA2',
 487                  *         because we already covered the range;
 488                  *    4.2) VMA2 and VMA2' covers different ranges, process
 489                  *         VMA2'.
 490                  */
 491                 if (mmap_lock_is_contended(curr_mm)) {
 492                         info->prev_vm_start = curr_vma->vm_start;
 493                         info->prev_vm_end = curr_vma->vm_end;
 494                         op = task_vma_iter_find_vma;
 495                         mmap_read_unlock(curr_mm);
 496                         if (mmap_read_lock_killable(curr_mm)) {
 497                                 mmput(curr_mm);
 498                                 goto finish;
 499                         }
 500                 } else {
 501                         op = task_vma_iter_next_vma;
 502                 }
 503         } else {
 504 again:
 505                 curr_task = task_seq_get_next(&info->common, &info->tid, true);
 506                 if (!curr_task) {
 507                         info->tid++;
 508                         goto finish;
 509                 }
 510
 511                 if (saved_tid != info->tid) {
 512                         /* new task, process the first vma */
 513                         op = task_vma_iter_first_vma;
 514                 } else {
 515                         /* Found the same tid, which means the user space
 516                          * finished data in previous buffer and read more.
 517                          * We dropped mmap_lock before returning to user
 518                          * space, so it is necessary to use find_vma() to
 519                          * find the next vma to process.
 520                          */
 521                         op = task_vma_iter_find_vma;
 522                 }
 523
 524                 curr_mm = get_task_mm(curr_task);
 525                 if (!curr_mm)
 526                         goto next_task;
 527
 528                 if (mmap_read_lock_killable(curr_mm)) {
 529                         mmput(curr_mm);
 530                         goto finish;
 531                 }
 532         }
 533
 534         switch (op) {
 535         case task_vma_iter_first_vma:
 536                 curr_vma = find_vma(curr_mm, 0);
 537                 break;
 538         case task_vma_iter_next_vma:
 539                 curr_vma = find_vma(curr_mm, curr_vma->vm_end);
 540                 break;
 541         case task_vma_iter_find_vma:
 542                 /* We dropped mmap_lock so it is necessary to use find_vma
 543                  * to find the next vma. This is similar to the  mechanism
 544                  * in show_smaps_rollup().
 545                  */
 546                 curr_vma = find_vma(curr_mm, info->prev_vm_end - 1);
 547                 /* case 1) and 4.2) above just use curr_vma */
 548
 549                 /* check for case 2) or case 4.1) above */
 550                 if (curr_vma &&
 551                     curr_vma->vm_start == info->prev_vm_start &&
 552                     curr_vma->vm_end == info->prev_vm_end)
 553                         curr_vma = find_vma(curr_mm, curr_vma->vm_end);
 554                 break;
 555         }
 556         if (!curr_vma) {
 557                 /* case 3) above, or case 2) 4.1) with vma->next == NULL */
 558                 mmap_read_unlock(curr_mm);
 559                 mmput(curr_mm);
 560                 goto next_task;
 561         }
 562         info->task = curr_task;
 563         info->vma = curr_vma;
 564         info->mm = curr_mm;
 565         return curr_vma;
 566
 567 next_task:
 568         if (info->common.type == BPF_TASK_ITER_TID)
 569                 goto finish;
 570
 571         put_task_struct(curr_task);
 572         info->task = NULL;
 573         info->mm = NULL;
 574         info->tid++;
 575         goto again;
 576
 577 finish:
 578         if (curr_task)
 579                 put_task_struct(curr_task);
 580         info->task = NULL;
 581         info->vma = NULL;
 582         info->mm = NULL;
 583         return NULL;
 584 }
 585
 586 static void *task_vma_seq_start(struct seq_file *seq, loff_t *pos)
 587 {
 588         struct bpf_iter_seq_task_vma_info *info = seq->private;
 589         struct vm_area_struct *vma;
 590
 591         vma = task_vma_seq_get_next(info);
 592         if (vma && *pos == 0)
 593                 ++*pos;
 594
 595         return vma;
 596 }
 597
 598 static void *task_vma_seq_next(struct seq_file *seq, void *v, loff_t *pos)
 599 {
 600         struct bpf_iter_seq_task_vma_info *info = seq->private;
 601
 602         ++*pos;
 603         return task_vma_seq_get_next(info);
 604 }
 605
 606 struct bpf_iter__task_vma {
 607         __bpf_md_ptr(struct bpf_iter_meta *, meta);
 608         __bpf_md_ptr(struct task_struct *, task);
 609         __bpf_md_ptr(struct vm_area_struct *, vma);
 610 };
 611
 612 DEFINE_BPF_ITER_FUNC(task_vma, struct bpf_iter_meta *meta,
 613                      struct task_struct *task, struct vm_area_struct *vma)
 614
 615 static int __task_vma_seq_show(struct seq_file *seq, bool in_stop)
 616 {
 617         struct bpf_iter_seq_task_vma_info *info = seq->private;
 618         struct bpf_iter__task_vma ctx;
 619         struct bpf_iter_meta meta;
 620         struct bpf_prog *prog;
 621
 622         meta.seq = seq;
 623         prog = bpf_iter_get_info(&meta, in_stop);
 624         if (!prog)
 625                 return 0;
 626
 627         ctx.meta = &meta;
 628         ctx.task = info->task;
 629         ctx.vma = info->vma;
 630         return bpf_iter_run_prog(prog, &ctx);
 631 }
 632
 633 static int task_vma_seq_show(struct seq_file *seq, void *v)
 634 {
 635         return __task_vma_seq_show(seq, false);
 636 }
 637
 638 static void task_vma_seq_stop(struct seq_file *seq, void *v)
 639 {
 640         struct bpf_iter_seq_task_vma_info *info = seq->private;
 641
 642         if (!v) {
 643                 (void)__task_vma_seq_show(seq, true);
 644         } else {
 645                 /* info->vma has not been seen by the BPF program. If the
 646                  * user space reads more, task_vma_seq_get_next should
 647                  * return this vma again. Set prev_vm_start to ~0UL,
 648                  * so that we don't skip the vma returned by the next
 649                  * find_vma() (case task_vma_iter_find_vma in
 650                  * task_vma_seq_get_next()).
 651                  */
 652                 info->prev_vm_start = ~0UL;
 653                 info->prev_vm_end = info->vma->vm_end;
 654                 mmap_read_unlock(info->mm);
 655                 mmput(info->mm);
 656                 info->mm = NULL;
 657                 put_task_struct(info->task);
 658                 info->task = NULL;
 659         }
 660 }
 661
 662 static const struct seq_operations task_vma_seq_ops = {
 663         .start  = task_vma_seq_start,
 664         .next   = task_vma_seq_next,
 665         .stop   = task_vma_seq_stop,
 666         .show   = task_vma_seq_show,
 667 };
 668
 669 static const struct bpf_iter_seq_info task_seq_info = {
 670         .seq_ops                = &task_seq_ops,
 671         .init_seq_private       = init_seq_pidns,
 672         .fini_seq_private       = fini_seq_pidns,
 673         .seq_priv_size          = sizeof(struct bpf_iter_seq_task_info),
 674 };
 675
 676 static int bpf_iter_fill_link_info(const struct bpf_iter_aux_info *aux, struct bpf_link_info *info)
 677 {
 678         switch (aux->task.type) {
 679         case BPF_TASK_ITER_TID:
 680                 info->iter.task.tid = aux->task.pid;
 681                 break;
 682         case BPF_TASK_ITER_TGID:
 683                 info->iter.task.pid = aux->task.pid;
 684                 break;
 685         default:
 686                 break;
 687         }
 688         return 0;
 689 }
 690
 691 static void bpf_iter_task_show_fdinfo(const struct bpf_iter_aux_info *aux, struct seq_file *seq)
 692 {
 693         seq_printf(seq, "task_type:\t%s\n", iter_task_type_names[aux->task.type]);
 694         if (aux->task.type == BPF_TASK_ITER_TID)
 695                 seq_printf(seq, "tid:\t%u\n", aux->task.pid);
 696         else if (aux->task.type == BPF_TASK_ITER_TGID)
 697                 seq_printf(seq, "pid:\t%u\n", aux->task.pid);
 698 }
 699
 700 static struct bpf_iter_reg task_reg_info = {
 701         .target                 = "task",
 702         .attach_target          = bpf_iter_attach_task,
 703         .feature                = BPF_ITER_RESCHED,
 704         .ctx_arg_info_size      = 1,
 705         .ctx_arg_info           = {
 706                 { offsetof(struct bpf_iter__task, task),
 707                   PTR_TO_BTF_ID_OR_NULL | PTR_TRUSTED },
 708         },
 709         .seq_info               = &task_seq_info,
 710         .fill_link_info         = bpf_iter_fill_link_info,
 711         .show_fdinfo            = bpf_iter_task_show_fdinfo,
 712 };
 713
 714 static const struct bpf_iter_seq_info task_file_seq_info = {
 715         .seq_ops                = &task_file_seq_ops,
 716         .init_seq_private       = init_seq_pidns,
 717         .fini_seq_private       = fini_seq_pidns,
 718         .seq_priv_size          = sizeof(struct bpf_iter_seq_task_file_info),
 719 };
 720
 721 static struct bpf_iter_reg task_file_reg_info = {
 722         .target                 = "task_file",
 723         .attach_target          = bpf_iter_attach_task,
 724         .feature                = BPF_ITER_RESCHED,
 725         .ctx_arg_info_size      = 2,
 726         .ctx_arg_info           = {
 727                 { offsetof(struct bpf_iter__task_file, task),
 728                   PTR_TO_BTF_ID_OR_NULL },
 729                 { offsetof(struct bpf_iter__task_file, file),
 730                   PTR_TO_BTF_ID_OR_NULL },
 731         },
 732         .seq_info               = &task_file_seq_info,
 733         .fill_link_info         = bpf_iter_fill_link_info,
 734         .show_fdinfo            = bpf_iter_task_show_fdinfo,
 735 };
 736
 737 static const struct bpf_iter_seq_info task_vma_seq_info = {
 738         .seq_ops                = &task_vma_seq_ops,
 739         .init_seq_private       = init_seq_pidns,
 740         .fini_seq_private       = fini_seq_pidns,
 741         .seq_priv_size          = sizeof(struct bpf_iter_seq_task_vma_info),
 742 };
 743
 744 static struct bpf_iter_reg task_vma_reg_info = {
 745         .target                 = "task_vma",
 746         .attach_target          = bpf_iter_attach_task,
 747         .feature                = BPF_ITER_RESCHED,
 748         .ctx_arg_info_size      = 2,
 749         .ctx_arg_info           = {
 750                 { offsetof(struct bpf_iter__task_vma, task),
 751                   PTR_TO_BTF_ID_OR_NULL },
 752                 { offsetof(struct bpf_iter__task_vma, vma),
 753                   PTR_TO_BTF_ID_OR_NULL },
 754         },
 755         .seq_info               = &task_vma_seq_info,
 756         .fill_link_info         = bpf_iter_fill_link_info,
 757         .show_fdinfo            = bpf_iter_task_show_fdinfo,
 758 };
 759
 760 BPF_CALL_5(bpf_find_vma, struct task_struct *, task, u64, start,
 761            bpf_callback_t, callback_fn, void *, callback_ctx, u64, flags)
 762 {
 763         struct mmap_unlock_irq_work *work = NULL;
 764         struct vm_area_struct *vma;
 765         bool irq_work_busy = false;
 766         struct mm_struct *mm;
 767         int ret = -ENOENT;
 768
 769         if (flags)
 770                 return -EINVAL;
 771
 772         if (!task)
 773                 return -ENOENT;
 774
 775         mm = task->mm;
 776         if (!mm)
 777                 return -ENOENT;
 778
 779         irq_work_busy = bpf_mmap_unlock_get_irq_work(&work);
 780
 781         if (irq_work_busy || !mmap_read_trylock(mm))
 782                 return -EBUSY;
 783
 784         vma = find_vma(mm, start);
 785
 786         if (vma && vma->vm_start <= start && vma->vm_end > start) {
 787                 callback_fn((u64)(long)task, (u64)(long)vma,
 788                             (u64)(long)callback_ctx, 0, 0);
 789                 ret = 0;
 790         }
 791         bpf_mmap_unlock_mm(work, mm);
 792         return ret;
 793 }
 794
 795 const struct bpf_func_proto bpf_find_vma_proto = {
 796         .func           = bpf_find_vma,
 797         .ret_type       = RET_INTEGER,
 798         .arg1_type      = ARG_PTR_TO_BTF_ID,
 799         .arg1_btf_id    = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
 800         .arg2_type      = ARG_ANYTHING,
 801         .arg3_type      = ARG_PTR_TO_FUNC,
 802         .arg4_type      = ARG_PTR_TO_STACK_OR_NULL,
 803         .arg5_type      = ARG_ANYTHING,
 804 };
 805
 806 struct bpf_iter_task_vma_kern_data {
 807         struct task_struct *task;
 808         struct mm_struct *mm;
 809         struct mmap_unlock_irq_work *work;
 810         struct vma_iterator vmi;
 811 };
 812
 813 struct bpf_iter_task_vma {
 814         /* opaque iterator state; having __u64 here allows to preserve correct
 815          * alignment requirements in vmlinux.h, generated from BTF
 816          */
 817         __u64 __opaque[1];
 818 } __attribute__((aligned(8)));
 819
 820 /* Non-opaque version of bpf_iter_task_vma */
 821 struct bpf_iter_task_vma_kern {
 822         struct bpf_iter_task_vma_kern_data *data;
 823 } __attribute__((aligned(8)));
 824
 825 __bpf_kfunc_start_defs();
 826
 827 __bpf_kfunc int bpf_iter_task_vma_new(struct bpf_iter_task_vma *it,
 828                                       struct task_struct *task, u64 addr)
 829 {
 830         struct bpf_iter_task_vma_kern *kit = (void *)it;
 831         bool irq_work_busy = false;
 832         int err;
 833
 834         BUILD_BUG_ON(sizeof(struct bpf_iter_task_vma_kern) != sizeof(struct bpf_iter_task_vma));
 835         BUILD_BUG_ON(__alignof__(struct bpf_iter_task_vma_kern) != __alignof__(struct bpf_iter_task_vma));
 836
 837         /* is_iter_reg_valid_uninit guarantees that kit hasn't been initialized
 838          * before, so non-NULL kit->data doesn't point to previously
 839          * bpf_mem_alloc'd bpf_iter_task_vma_kern_data
 840          */
 841         kit->data = bpf_mem_alloc(&bpf_global_ma, sizeof(struct bpf_iter_task_vma_kern_data));
 842         if (!kit->data)
 843                 return -ENOMEM;
 844
 845         kit->data->task = get_task_struct(task);
 846         kit->data->mm = task->mm;
 847         if (!kit->data->mm) {
 848                 err = -ENOENT;
 849                 goto err_cleanup_iter;
 850         }
 851
 852         /* kit->data->work == NULL is valid after bpf_mmap_unlock_get_irq_work */
 853         irq_work_busy = bpf_mmap_unlock_get_irq_work(&kit->data->work);
 854         if (irq_work_busy || !mmap_read_trylock(kit->data->mm)) {
 855                 err = -EBUSY;
 856                 goto err_cleanup_iter;
 857         }
 858
 859         vma_iter_init(&kit->data->vmi, kit->data->mm, addr);
 860         return 0;
 861
 862 err_cleanup_iter:
 863         if (kit->data->task)
 864                 put_task_struct(kit->data->task);
 865         bpf_mem_free(&bpf_global_ma, kit->data);
 866         /* NULL kit->data signals failed bpf_iter_task_vma initialization */
 867         kit->data = NULL;
 868         return err;
 869 }
 870
 871 __bpf_kfunc struct vm_area_struct *bpf_iter_task_vma_next(struct bpf_iter_task_vma *it)
 872 {
 873         struct bpf_iter_task_vma_kern *kit = (void *)it;
 874
 875         if (!kit->data) /* bpf_iter_task_vma_new failed */
 876                 return NULL;
 877         return vma_next(&kit->data->vmi);
 878 }
 879
 880 __bpf_kfunc void bpf_iter_task_vma_destroy(struct bpf_iter_task_vma *it)
 881 {
 882         struct bpf_iter_task_vma_kern *kit = (void *)it;
 883
 884         if (kit->data) {
 885                 bpf_mmap_unlock_mm(kit->data->work, kit->data->mm);
 886                 put_task_struct(kit->data->task);
 887                 bpf_mem_free(&bpf_global_ma, kit->data);
 888         }
 889 }
 890
 891 __bpf_kfunc_end_defs();
 892
 893 #ifdef CONFIG_CGROUPS
 894
 895 struct bpf_iter_css_task {
 896         __u64 __opaque[1];
 897 } __attribute__((aligned(8)));
 898
 899 struct bpf_iter_css_task_kern {
 900         struct css_task_iter *css_it;
 901 } __attribute__((aligned(8)));
 902
 903 __bpf_kfunc_start_defs();
 904
 905 __bpf_kfunc int bpf_iter_css_task_new(struct bpf_iter_css_task *it,
 906                 struct cgroup_subsys_state *css, unsigned int flags)
 907 {
 908         struct bpf_iter_css_task_kern *kit = (void *)it;
 909
 910         BUILD_BUG_ON(sizeof(struct bpf_iter_css_task_kern) != sizeof(struct bpf_iter_css_task));
 911         BUILD_BUG_ON(__alignof__(struct bpf_iter_css_task_kern) !=
 912                                         __alignof__(struct bpf_iter_css_task));
 913         kit->css_it = NULL;
 914         switch (flags) {
 915         case CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED:
 916         case CSS_TASK_ITER_PROCS:
 917         case 0:
 918                 break;
 919         default:
 920                 return -EINVAL;
 921         }
 922
 923         kit->css_it = bpf_mem_alloc(&bpf_global_ma, sizeof(struct css_task_iter));
 924         if (!kit->css_it)
 925                 return -ENOMEM;
 926         css_task_iter_start(css, flags, kit->css_it);
 927         return 0;
 928 }
 929
 930 __bpf_kfunc struct task_struct *bpf_iter_css_task_next(struct bpf_iter_css_task *it)
 931 {
 932         struct bpf_iter_css_task_kern *kit = (void *)it;
 933
 934         if (!kit->css_it)
 935                 return NULL;
 936         return css_task_iter_next(kit->css_it);
 937 }
 938
 939 __bpf_kfunc void bpf_iter_css_task_destroy(struct bpf_iter_css_task *it)
 940 {
 941         struct bpf_iter_css_task_kern *kit = (void *)it;
 942
 943         if (!kit->css_it)
 944                 return;
 945         css_task_iter_end(kit->css_it);
 946         bpf_mem_free(&bpf_global_ma, kit->css_it);
 947 }
 948
 949 __bpf_kfunc_end_defs();
 950
 951 #endif /* CONFIG_CGROUPS */
 952
 953 struct bpf_iter_task {
 954         __u64 __opaque[3];
 955 } __attribute__((aligned(8)));
 956
 957 struct bpf_iter_task_kern {
 958         struct task_struct *task;
 959         struct task_struct *pos;
 960         unsigned int flags;
 961 } __attribute__((aligned(8)));
 962
 963 enum {
 964         /* all process in the system */
 965         BPF_TASK_ITER_ALL_PROCS,
 966         /* all threads in the system */
 967         BPF_TASK_ITER_ALL_THREADS,
 968         /* all threads of a specific process */
 969         BPF_TASK_ITER_PROC_THREADS
 970 };
 971
 972 __bpf_kfunc_start_defs();
 973
 974 __bpf_kfunc int bpf_iter_task_new(struct bpf_iter_task *it,
 975                 struct task_struct *task__nullable, unsigned int flags)
 976 {
 977         struct bpf_iter_task_kern *kit = (void *)it;
 978
 979         BUILD_BUG_ON(sizeof(struct bpf_iter_task_kern) > sizeof(struct bpf_iter_task));
 980         BUILD_BUG_ON(__alignof__(struct bpf_iter_task_kern) !=
 981                                         __alignof__(struct bpf_iter_task));
 982
 983         kit->task = kit->pos = NULL;
 984         switch (flags) {
 985         case BPF_TASK_ITER_ALL_THREADS:
 986         case BPF_TASK_ITER_ALL_PROCS:
 987                 break;
 988         case BPF_TASK_ITER_PROC_THREADS:
 989                 if (!task__nullable)
 990                         return -EINVAL;
 991                 break;
 992         default:
 993                 return -EINVAL;
 994         }
 995
 996         if (flags == BPF_TASK_ITER_PROC_THREADS)
 997                 kit->task = task__nullable;
 998         else
 999                 kit->task = &init_task;
1000         kit->pos = kit->task;
1001         kit->flags = flags;
1002         return 0;
1003 }
1004
1005 __bpf_kfunc struct task_struct *bpf_iter_task_next(struct bpf_iter_task *it)
1006 {
1007         struct bpf_iter_task_kern *kit = (void *)it;
1008         struct task_struct *pos;
1009         unsigned int flags;
1010
1011         flags = kit->flags;
1012         pos = kit->pos;
1013
1014         if (!pos)
1015                 return pos;
1016
1017         if (flags == BPF_TASK_ITER_ALL_PROCS)
1018                 goto get_next_task;
1019
1020         kit->pos = next_thread(kit->pos);
1021         if (kit->pos == kit->task) {
1022                 if (flags == BPF_TASK_ITER_PROC_THREADS) {
1023                         kit->pos = NULL;
1024                         return pos;
1025                 }
1026         } else
1027                 return pos;
1028
1029 get_next_task:
1030         kit->pos = next_task(kit->pos);
1031         kit->task = kit->pos;
1032         if (kit->pos == &init_task)
1033                 kit->pos = NULL;
1034
1035         return pos;
1036 }
1037
1038 __bpf_kfunc void bpf_iter_task_destroy(struct bpf_iter_task *it)
1039 {
1040 }
1041
1042 __bpf_kfunc_end_defs();
1043
1044 DEFINE_PER_CPU(struct mmap_unlock_irq_work, mmap_unlock_work);
1045
1046 static void do_mmap_read_unlock(struct irq_work *entry)
1047 {
1048         struct mmap_unlock_irq_work *work;
1049
1050         if (WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_RT)))
1051                 return;
1052
1053         work = container_of(entry, struct mmap_unlock_irq_work, irq_work);
1054         mmap_read_unlock_non_owner(work->mm);
1055 }
1056
1057 static int __init task_iter_init(void)
1058 {
1059         struct mmap_unlock_irq_work *work;
1060         int ret, cpu;
1061
1062         for_each_possible_cpu(cpu) {
1063                 work = per_cpu_ptr(&mmap_unlock_work, cpu);
1064                 init_irq_work(&work->irq_work, do_mmap_read_unlock);
1065         }
1066
1067         task_reg_info.ctx_arg_info[0].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_TASK];
1068         ret = bpf_iter_reg_target(&task_reg_info);
1069         if (ret)
1070                 return ret;
1071
1072         task_file_reg_info.ctx_arg_info[0].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_TASK];
1073         task_file_reg_info.ctx_arg_info[1].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_FILE];
1074         ret =  bpf_iter_reg_target(&task_file_reg_info);
1075         if (ret)
1076                 return ret;
1077
1078         task_vma_reg_info.ctx_arg_info[0].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_TASK];
1079         task_vma_reg_info.ctx_arg_info[1].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA];
1080         return bpf_iter_reg_target(&task_vma_reg_info);
1081 }
1082 late_initcall(task_iter_init);