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/btf_ids.h>
  11 #include "mmap_unlock_work.h"
  12
  13 struct bpf_iter_seq_task_common {
  14         struct pid_namespace *ns;
  15 };
  16
  17 struct bpf_iter_seq_task_info {
  18         /* The first field must be struct bpf_iter_seq_task_common.
  19          * this is assumed by {init, fini}_seq_pidns() callback functions.
  20          */
  21         struct bpf_iter_seq_task_common common;
  22         u32 tid;
  23 };
  24
  25 static struct task_struct *task_seq_get_next(struct pid_namespace *ns,
  26                                              u32 *tid,
  27                                              bool skip_if_dup_files)
  28 {
  29         struct task_struct *task = NULL;
  30         struct pid *pid;
  31
  32         rcu_read_lock();
  33 retry:
  34         pid = find_ge_pid(*tid, ns);
  35         if (pid) {
  36                 *tid = pid_nr_ns(pid, ns);
  37                 task = get_pid_task(pid, PIDTYPE_PID);
  38                 if (!task) {
  39                         ++*tid;
  40                         goto retry;
  41                 } else if (skip_if_dup_files && !thread_group_leader(task) &&
  42                            task->files == task->group_leader->files) {
  43                         put_task_struct(task);
  44                         task = NULL;
  45                         ++*tid;
  46                         goto retry;
  47                 }
  48         }
  49         rcu_read_unlock();
  50
  51         return task;
  52 }
  53
  54 static void *task_seq_start(struct seq_file *seq, loff_t *pos)
  55 {
  56         struct bpf_iter_seq_task_info *info = seq->private;
  57         struct task_struct *task;
  58
  59         task = task_seq_get_next(info->common.ns, &info->tid, false);
  60         if (!task)
  61                 return NULL;
  62
  63         if (*pos == 0)
  64                 ++*pos;
  65         return task;
  66 }
  67
  68 static void *task_seq_next(struct seq_file *seq, void *v, loff_t *pos)
  69 {
  70         struct bpf_iter_seq_task_info *info = seq->private;
  71         struct task_struct *task;
  72
  73         ++*pos;
  74         ++info->tid;
  75         put_task_struct((struct task_struct *)v);
  76         task = task_seq_get_next(info->common.ns, &info->tid, false);
  77         if (!task)
  78                 return NULL;
  79
  80         return task;
  81 }
  82
  83 struct bpf_iter__task {
  84         __bpf_md_ptr(struct bpf_iter_meta *, meta);
  85         __bpf_md_ptr(struct task_struct *, task);
  86 };
  87
  88 DEFINE_BPF_ITER_FUNC(task, struct bpf_iter_meta *meta, struct task_struct *task)
  89
  90 static int __task_seq_show(struct seq_file *seq, struct task_struct *task,
  91                            bool in_stop)
  92 {
  93         struct bpf_iter_meta meta;
  94         struct bpf_iter__task ctx;
  95         struct bpf_prog *prog;
  96
  97         meta.seq = seq;
  98         prog = bpf_iter_get_info(&meta, in_stop);
  99         if (!prog)
 100                 return 0;
 101
 102         ctx.meta = &meta;
 103         ctx.task = task;
 104         return bpf_iter_run_prog(prog, &ctx);
 105 }
 106
 107 static int task_seq_show(struct seq_file *seq, void *v)
 108 {
 109         return __task_seq_show(seq, v, false);
 110 }
 111
 112 static void task_seq_stop(struct seq_file *seq, void *v)
 113 {
 114         if (!v)
 115                 (void)__task_seq_show(seq, v, true);
 116         else
 117                 put_task_struct((struct task_struct *)v);
 118 }
 119
 120 static const struct seq_operations task_seq_ops = {
 121         .start  = task_seq_start,
 122         .next   = task_seq_next,
 123         .stop   = task_seq_stop,
 124         .show   = task_seq_show,
 125 };
 126
 127 struct bpf_iter_seq_task_file_info {
 128         /* The first field must be struct bpf_iter_seq_task_common.
 129          * this is assumed by {init, fini}_seq_pidns() callback functions.
 130          */
 131         struct bpf_iter_seq_task_common common;
 132         struct task_struct *task;
 133         u32 tid;
 134         u32 fd;
 135 };
 136
 137 static struct file *
 138 task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info)
 139 {
 140         struct pid_namespace *ns = info->common.ns;
 141         u32 curr_tid = info->tid;
 142         struct task_struct *curr_task;
 143         unsigned int curr_fd = info->fd;
 144
 145         /* If this function returns a non-NULL file object,
 146          * it held a reference to the task/file.
 147          * Otherwise, it does not hold any reference.
 148          */
 149 again:
 150         if (info->task) {
 151                 curr_task = info->task;
 152                 curr_fd = info->fd;
 153         } else {
 154                 curr_task = task_seq_get_next(ns, &curr_tid, true);
 155                 if (!curr_task) {
 156                         info->task = NULL;
 157                         info->tid = curr_tid;
 158                         return NULL;
 159                 }
 160
 161                 /* set info->task and info->tid */
 162                 info->task = curr_task;
 163                 if (curr_tid == info->tid) {
 164                         curr_fd = info->fd;
 165                 } else {
 166                         info->tid = curr_tid;
 167                         curr_fd = 0;
 168                 }
 169         }
 170
 171         rcu_read_lock();
 172         for (;; curr_fd++) {
 173                 struct file *f;
 174                 f = task_lookup_next_fd_rcu(curr_task, &curr_fd);
 175                 if (!f)
 176                         break;
 177                 if (!get_file_rcu(f))
 178                         continue;
 179
 180                 /* set info->fd */
 181                 info->fd = curr_fd;
 182                 rcu_read_unlock();
 183                 return f;
 184         }
 185
 186         /* the current task is done, go to the next task */
 187         rcu_read_unlock();
 188         put_task_struct(curr_task);
 189         info->task = NULL;
 190         info->fd = 0;
 191         curr_tid = ++(info->tid);
 192         goto again;
 193 }
 194
 195 static void *task_file_seq_start(struct seq_file *seq, loff_t *pos)
 196 {
 197         struct bpf_iter_seq_task_file_info *info = seq->private;
 198         struct file *file;
 199
 200         info->task = NULL;
 201         file = task_file_seq_get_next(info);
 202         if (file && *pos == 0)
 203                 ++*pos;
 204
 205         return file;
 206 }
 207
 208 static void *task_file_seq_next(struct seq_file *seq, void *v, loff_t *pos)
 209 {
 210         struct bpf_iter_seq_task_file_info *info = seq->private;
 211
 212         ++*pos;
 213         ++info->fd;
 214         fput((struct file *)v);
 215         return task_file_seq_get_next(info);
 216 }
 217
 218 struct bpf_iter__task_file {
 219         __bpf_md_ptr(struct bpf_iter_meta *, meta);
 220         __bpf_md_ptr(struct task_struct *, task);
 221         u32 fd __aligned(8);
 222         __bpf_md_ptr(struct file *, file);
 223 };
 224
 225 DEFINE_BPF_ITER_FUNC(task_file, struct bpf_iter_meta *meta,
 226                      struct task_struct *task, u32 fd,
 227                      struct file *file)
 228
 229 static int __task_file_seq_show(struct seq_file *seq, struct file *file,
 230                                 bool in_stop)
 231 {
 232         struct bpf_iter_seq_task_file_info *info = seq->private;
 233         struct bpf_iter__task_file ctx;
 234         struct bpf_iter_meta meta;
 235         struct bpf_prog *prog;
 236
 237         meta.seq = seq;
 238         prog = bpf_iter_get_info(&meta, in_stop);
 239         if (!prog)
 240                 return 0;
 241
 242         ctx.meta = &meta;
 243         ctx.task = info->task;
 244         ctx.fd = info->fd;
 245         ctx.file = file;
 246         return bpf_iter_run_prog(prog, &ctx);
 247 }
 248
 249 static int task_file_seq_show(struct seq_file *seq, void *v)
 250 {
 251         return __task_file_seq_show(seq, v, false);
 252 }
 253
 254 static void task_file_seq_stop(struct seq_file *seq, void *v)
 255 {
 256         struct bpf_iter_seq_task_file_info *info = seq->private;
 257
 258         if (!v) {
 259                 (void)__task_file_seq_show(seq, v, true);
 260         } else {
 261                 fput((struct file *)v);
 262                 put_task_struct(info->task);
 263                 info->task = NULL;
 264         }
 265 }
 266
 267 static int init_seq_pidns(void *priv_data, struct bpf_iter_aux_info *aux)
 268 {
 269         struct bpf_iter_seq_task_common *common = priv_data;
 270
 271         common->ns = get_pid_ns(task_active_pid_ns(current));
 272         return 0;
 273 }
 274
 275 static void fini_seq_pidns(void *priv_data)
 276 {
 277         struct bpf_iter_seq_task_common *common = priv_data;
 278
 279         put_pid_ns(common->ns);
 280 }
 281
 282 static const struct seq_operations task_file_seq_ops = {
 283         .start  = task_file_seq_start,
 284         .next   = task_file_seq_next,
 285         .stop   = task_file_seq_stop,
 286         .show   = task_file_seq_show,
 287 };
 288
 289 struct bpf_iter_seq_task_vma_info {
 290         /* The first field must be struct bpf_iter_seq_task_common.
 291          * this is assumed by {init, fini}_seq_pidns() callback functions.
 292          */
 293         struct bpf_iter_seq_task_common common;
 294         struct task_struct *task;
 295         struct vm_area_struct *vma;
 296         u32 tid;
 297         unsigned long prev_vm_start;
 298         unsigned long prev_vm_end;
 299 };
 300
 301 enum bpf_task_vma_iter_find_op {
 302         task_vma_iter_first_vma,   /* use mm->mmap */
 303         task_vma_iter_next_vma,    /* use curr_vma->vm_next */
 304         task_vma_iter_find_vma,    /* use find_vma() to find next vma */
 305 };
 306
 307 static struct vm_area_struct *
 308 task_vma_seq_get_next(struct bpf_iter_seq_task_vma_info *info)
 309 {
 310         struct pid_namespace *ns = info->common.ns;
 311         enum bpf_task_vma_iter_find_op op;
 312         struct vm_area_struct *curr_vma;
 313         struct task_struct *curr_task;
 314         u32 curr_tid = info->tid;
 315
 316         /* If this function returns a non-NULL vma, it holds a reference to
 317          * the task_struct, and holds read lock on vma->mm->mmap_lock.
 318          * If this function returns NULL, it does not hold any reference or
 319          * lock.
 320          */
 321         if (info->task) {
 322                 curr_task = info->task;
 323                 curr_vma = info->vma;
 324                 /* In case of lock contention, drop mmap_lock to unblock
 325                  * the writer.
 326                  *
 327                  * After relock, call find(mm, prev_vm_end - 1) to find
 328                  * new vma to process.
 329                  *
 330                  *   +------+------+-----------+
 331                  *   | VMA1 | VMA2 | VMA3      |
 332                  *   +------+------+-----------+
 333                  *   |      |      |           |
 334                  *  4k     8k     16k         400k
 335                  *
 336                  * For example, curr_vma == VMA2. Before unlock, we set
 337                  *
 338                  *    prev_vm_start = 8k
 339                  *    prev_vm_end   = 16k
 340                  *
 341                  * There are a few cases:
 342                  *
 343                  * 1) VMA2 is freed, but VMA3 exists.
 344                  *
 345                  *    find_vma() will return VMA3, just process VMA3.
 346                  *
 347                  * 2) VMA2 still exists.
 348                  *
 349                  *    find_vma() will return VMA2, process VMA2->next.
 350                  *
 351                  * 3) no more vma in this mm.
 352                  *
 353                  *    Process the next task.
 354                  *
 355                  * 4) find_vma() returns a different vma, VMA2'.
 356                  *
 357                  *    4.1) If VMA2 covers same range as VMA2', skip VMA2',
 358                  *         because we already covered the range;
 359                  *    4.2) VMA2 and VMA2' covers different ranges, process
 360                  *         VMA2'.
 361                  */
 362                 if (mmap_lock_is_contended(curr_task->mm)) {
 363                         info->prev_vm_start = curr_vma->vm_start;
 364                         info->prev_vm_end = curr_vma->vm_end;
 365                         op = task_vma_iter_find_vma;
 366                         mmap_read_unlock(curr_task->mm);
 367                         if (mmap_read_lock_killable(curr_task->mm))
 368                                 goto finish;
 369                 } else {
 370                         op = task_vma_iter_next_vma;
 371                 }
 372         } else {
 373 again:
 374                 curr_task = task_seq_get_next(ns, &curr_tid, true);
 375                 if (!curr_task) {
 376                         info->tid = curr_tid + 1;
 377                         goto finish;
 378                 }
 379
 380                 if (curr_tid != info->tid) {
 381                         info->tid = curr_tid;
 382                         /* new task, process the first vma */
 383                         op = task_vma_iter_first_vma;
 384                 } else {
 385                         /* Found the same tid, which means the user space
 386                          * finished data in previous buffer and read more.
 387                          * We dropped mmap_lock before returning to user
 388                          * space, so it is necessary to use find_vma() to
 389                          * find the next vma to process.
 390                          */
 391                         op = task_vma_iter_find_vma;
 392                 }
 393
 394                 if (!curr_task->mm)
 395                         goto next_task;
 396
 397                 if (mmap_read_lock_killable(curr_task->mm))
 398                         goto finish;
 399         }
 400
 401         switch (op) {
 402         case task_vma_iter_first_vma:
 403                 curr_vma = curr_task->mm->mmap;
 404                 break;
 405         case task_vma_iter_next_vma:
 406                 curr_vma = curr_vma->vm_next;
 407                 break;
 408         case task_vma_iter_find_vma:
 409                 /* We dropped mmap_lock so it is necessary to use find_vma
 410                  * to find the next vma. This is similar to the  mechanism
 411                  * in show_smaps_rollup().
 412                  */
 413                 curr_vma = find_vma(curr_task->mm, info->prev_vm_end - 1);
 414                 /* case 1) and 4.2) above just use curr_vma */
 415
 416                 /* check for case 2) or case 4.1) above */
 417                 if (curr_vma &&
 418                     curr_vma->vm_start == info->prev_vm_start &&
 419                     curr_vma->vm_end == info->prev_vm_end)
 420                         curr_vma = curr_vma->vm_next;
 421                 break;
 422         }
 423         if (!curr_vma) {
 424                 /* case 3) above, or case 2) 4.1) with vma->next == NULL */
 425                 mmap_read_unlock(curr_task->mm);
 426                 goto next_task;
 427         }
 428         info->task = curr_task;
 429         info->vma = curr_vma;
 430         return curr_vma;
 431
 432 next_task:
 433         put_task_struct(curr_task);
 434         info->task = NULL;
 435         curr_tid++;
 436         goto again;
 437
 438 finish:
 439         if (curr_task)
 440                 put_task_struct(curr_task);
 441         info->task = NULL;
 442         info->vma = NULL;
 443         return NULL;
 444 }
 445
 446 static void *task_vma_seq_start(struct seq_file *seq, loff_t *pos)
 447 {
 448         struct bpf_iter_seq_task_vma_info *info = seq->private;
 449         struct vm_area_struct *vma;
 450
 451         vma = task_vma_seq_get_next(info);
 452         if (vma && *pos == 0)
 453                 ++*pos;
 454
 455         return vma;
 456 }
 457
 458 static void *task_vma_seq_next(struct seq_file *seq, void *v, loff_t *pos)
 459 {
 460         struct bpf_iter_seq_task_vma_info *info = seq->private;
 461
 462         ++*pos;
 463         return task_vma_seq_get_next(info);
 464 }
 465
 466 struct bpf_iter__task_vma {
 467         __bpf_md_ptr(struct bpf_iter_meta *, meta);
 468         __bpf_md_ptr(struct task_struct *, task);
 469         __bpf_md_ptr(struct vm_area_struct *, vma);
 470 };
 471
 472 DEFINE_BPF_ITER_FUNC(task_vma, struct bpf_iter_meta *meta,
 473                      struct task_struct *task, struct vm_area_struct *vma)
 474
 475 static int __task_vma_seq_show(struct seq_file *seq, bool in_stop)
 476 {
 477         struct bpf_iter_seq_task_vma_info *info = seq->private;
 478         struct bpf_iter__task_vma ctx;
 479         struct bpf_iter_meta meta;
 480         struct bpf_prog *prog;
 481
 482         meta.seq = seq;
 483         prog = bpf_iter_get_info(&meta, in_stop);
 484         if (!prog)
 485                 return 0;
 486
 487         ctx.meta = &meta;
 488         ctx.task = info->task;
 489         ctx.vma = info->vma;
 490         return bpf_iter_run_prog(prog, &ctx);
 491 }
 492
 493 static int task_vma_seq_show(struct seq_file *seq, void *v)
 494 {
 495         return __task_vma_seq_show(seq, false);
 496 }
 497
 498 static void task_vma_seq_stop(struct seq_file *seq, void *v)
 499 {
 500         struct bpf_iter_seq_task_vma_info *info = seq->private;
 501
 502         if (!v) {
 503                 (void)__task_vma_seq_show(seq, true);
 504         } else {
 505                 /* info->vma has not been seen by the BPF program. If the
 506                  * user space reads more, task_vma_seq_get_next should
 507                  * return this vma again. Set prev_vm_start to ~0UL,
 508                  * so that we don't skip the vma returned by the next
 509                  * find_vma() (case task_vma_iter_find_vma in
 510                  * task_vma_seq_get_next()).
 511                  */
 512                 info->prev_vm_start = ~0UL;
 513                 info->prev_vm_end = info->vma->vm_end;
 514                 mmap_read_unlock(info->task->mm);
 515                 put_task_struct(info->task);
 516                 info->task = NULL;
 517         }
 518 }
 519
 520 static const struct seq_operations task_vma_seq_ops = {
 521         .start  = task_vma_seq_start,
 522         .next   = task_vma_seq_next,
 523         .stop   = task_vma_seq_stop,
 524         .show   = task_vma_seq_show,
 525 };
 526
 527 static const struct bpf_iter_seq_info task_seq_info = {
 528         .seq_ops                = &task_seq_ops,
 529         .init_seq_private       = init_seq_pidns,
 530         .fini_seq_private       = fini_seq_pidns,
 531         .seq_priv_size          = sizeof(struct bpf_iter_seq_task_info),
 532 };
 533
 534 static struct bpf_iter_reg task_reg_info = {
 535         .target                 = "task",
 536         .feature                = BPF_ITER_RESCHED,
 537         .ctx_arg_info_size      = 1,
 538         .ctx_arg_info           = {
 539                 { offsetof(struct bpf_iter__task, task),
 540                   PTR_TO_BTF_ID_OR_NULL },
 541         },
 542         .seq_info               = &task_seq_info,
 543 };
 544
 545 static const struct bpf_iter_seq_info task_file_seq_info = {
 546         .seq_ops                = &task_file_seq_ops,
 547         .init_seq_private       = init_seq_pidns,
 548         .fini_seq_private       = fini_seq_pidns,
 549         .seq_priv_size          = sizeof(struct bpf_iter_seq_task_file_info),
 550 };
 551
 552 static struct bpf_iter_reg task_file_reg_info = {
 553         .target                 = "task_file",
 554         .feature                = BPF_ITER_RESCHED,
 555         .ctx_arg_info_size      = 2,
 556         .ctx_arg_info           = {
 557                 { offsetof(struct bpf_iter__task_file, task),
 558                   PTR_TO_BTF_ID_OR_NULL },
 559                 { offsetof(struct bpf_iter__task_file, file),
 560                   PTR_TO_BTF_ID_OR_NULL },
 561         },
 562         .seq_info               = &task_file_seq_info,
 563 };
 564
 565 static const struct bpf_iter_seq_info task_vma_seq_info = {
 566         .seq_ops                = &task_vma_seq_ops,
 567         .init_seq_private       = init_seq_pidns,
 568         .fini_seq_private       = fini_seq_pidns,
 569         .seq_priv_size          = sizeof(struct bpf_iter_seq_task_vma_info),
 570 };
 571
 572 static struct bpf_iter_reg task_vma_reg_info = {
 573         .target                 = "task_vma",
 574         .feature                = BPF_ITER_RESCHED,
 575         .ctx_arg_info_size      = 2,
 576         .ctx_arg_info           = {
 577                 { offsetof(struct bpf_iter__task_vma, task),
 578                   PTR_TO_BTF_ID_OR_NULL },
 579                 { offsetof(struct bpf_iter__task_vma, vma),
 580                   PTR_TO_BTF_ID_OR_NULL },
 581         },
 582         .seq_info               = &task_vma_seq_info,
 583 };
 584
 585 BPF_CALL_5(bpf_find_vma, struct task_struct *, task, u64, start,
 586            bpf_callback_t, callback_fn, void *, callback_ctx, u64, flags)
 587 {
 588         struct mmap_unlock_irq_work *work = NULL;
 589         struct vm_area_struct *vma;
 590         bool irq_work_busy = false;
 591         struct mm_struct *mm;
 592         int ret = -ENOENT;
 593
 594         if (flags)
 595                 return -EINVAL;
 596
 597         if (!task)
 598                 return -ENOENT;
 599
 600         mm = task->mm;
 601         if (!mm)
 602                 return -ENOENT;
 603
 604         irq_work_busy = bpf_mmap_unlock_get_irq_work(&work);
 605
 606         if (irq_work_busy || !mmap_read_trylock(mm))
 607                 return -EBUSY;
 608
 609         vma = find_vma(mm, start);
 610
 611         if (vma && vma->vm_start <= start && vma->vm_end > start) {
 612                 callback_fn((u64)(long)task, (u64)(long)vma,
 613                             (u64)(long)callback_ctx, 0, 0);
 614                 ret = 0;
 615         }
 616         bpf_mmap_unlock_mm(work, mm);
 617         return ret;
 618 }
 619
 620 const struct bpf_func_proto bpf_find_vma_proto = {
 621         .func           = bpf_find_vma,
 622         .ret_type       = RET_INTEGER,
 623         .arg1_type      = ARG_PTR_TO_BTF_ID,
 624         .arg1_btf_id    = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
 625         .arg2_type      = ARG_ANYTHING,
 626         .arg3_type      = ARG_PTR_TO_FUNC,
 627         .arg4_type      = ARG_PTR_TO_STACK_OR_NULL,
 628         .arg5_type      = ARG_ANYTHING,
 629 };
 630
 631 DEFINE_PER_CPU(struct mmap_unlock_irq_work, mmap_unlock_work);
 632
 633 static void do_mmap_read_unlock(struct irq_work *entry)
 634 {
 635         struct mmap_unlock_irq_work *work;
 636
 637         if (WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_RT)))
 638                 return;
 639
 640         work = container_of(entry, struct mmap_unlock_irq_work, irq_work);
 641         mmap_read_unlock_non_owner(work->mm);
 642 }
 643
 644 static int __init task_iter_init(void)
 645 {
 646         struct mmap_unlock_irq_work *work;
 647         int ret, cpu;
 648
 649         for_each_possible_cpu(cpu) {
 650                 work = per_cpu_ptr(&mmap_unlock_work, cpu);
 651                 init_irq_work(&work->irq_work, do_mmap_read_unlock);
 652         }
 653
 654         task_reg_info.ctx_arg_info[0].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_TASK];
 655         ret = bpf_iter_reg_target(&task_reg_info);
 656         if (ret)
 657                 return ret;
 658
 659         task_file_reg_info.ctx_arg_info[0].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_TASK];
 660         task_file_reg_info.ctx_arg_info[1].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_FILE];
 661         ret =  bpf_iter_reg_target(&task_file_reg_info);
 662         if (ret)
 663                 return ret;
 664
 665         task_vma_reg_info.ctx_arg_info[0].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_TASK];
 666         task_vma_reg_info.ctx_arg_info[1].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA];
 667         return bpf_iter_reg_target(&task_vma_reg_info);
 668 }
 669 late_initcall(task_iter_init);