From 898720fdb7cd6db936cf4812171caea8b7143817 Mon Sep 17 00:00:00 2001
From: Frances Wingerter <fw@immunant.com>
Date: Mon, 22 Jan 2024 11:41:31 -0500
Subject: [PATCH] runtime: trace madvise syscall

madvise(MADV_DONTNEED) is equivalent to zeroing memory; we cannot allow this to be performed cross-compartment

because some other 'advices' may also have similar effects, for now we simply forbid all cross-compartment madvise(advice) calls regardless of the value of advice
---
 runtime/mmap_event.c       |  2 ++
 runtime/mmap_event.h       |  8 ++++++++
 runtime/seccomp_filter.c   |  1 +
 runtime/track_memory_map.c | 22 ++++++++++++++++++++++
 4 files changed, 33 insertions(+)

diff --git a/runtime/mmap_event.c b/runtime/mmap_event.c
index 070d11ac2..6b57162d3 100644
--- a/runtime/mmap_event.c
+++ b/runtime/mmap_event.c
@@ -10,6 +10,8 @@ enum mmap_event event_from_syscall(uint64_t rax) {
     return EVENT_MUNMAP;
   case __NR_mremap:
     return EVENT_MREMAP;
+  case __NR_madvise:
+    return EVENT_MADVISE;
   case __NR_mprotect:
     return EVENT_MPROTECT;
   case __NR_pkey_mprotect:
diff --git a/runtime/mmap_event.h b/runtime/mmap_event.h
index 0530c63f2..892c73bac 100644
--- a/runtime/mmap_event.h
+++ b/runtime/mmap_event.h
@@ -22,6 +22,11 @@ struct mremap_info {
   unsigned char pkey;
 };
 
+struct madvise_info {
+  struct range range;
+  unsigned char pkey;
+};
+
 struct mprotect_info {
   struct range range;
   int prot;
@@ -39,6 +44,7 @@ union event_info {
   struct mmap_info mmap;
   struct munmap_info munmap;
   struct mremap_info mremap;
+  struct madvise_info madvise;
   struct mprotect_info mprotect;
   struct pkey_mprotect_info pkey_mprotect;
 };
@@ -47,6 +53,7 @@ enum mmap_event {
   EVENT_MMAP,
   EVENT_MUNMAP,
   EVENT_MREMAP,
+  EVENT_MADVISE,
   EVENT_MPROTECT,
   EVENT_PKEY_MPROTECT,
   EVENT_CLONE,
@@ -58,6 +65,7 @@ static const char *event_names[] = {
     "MMAP",
     "MUNMAP",
     "MREMAP",
+    "MADVISE",
     "MPROTECT",
     "PKEY_MPROTECT",
     "CLONE",
diff --git a/runtime/seccomp_filter.c b/runtime/seccomp_filter.c
index 951cb547e..605e1faef 100644
--- a/runtime/seccomp_filter.c
+++ b/runtime/seccomp_filter.c
@@ -34,6 +34,7 @@ struct sock_filter ia2_filter[] = {
     BPF_SYSCALL_POLICY(mprotect, TRACE),
     BPF_SYSCALL_POLICY(mremap, TRACE),
     BPF_SYSCALL_POLICY(munmap, TRACE),
+    BPF_SYSCALL_POLICY(madvise, TRACE),
     /* pkey syscalls */
     BPF_SYSCALL_POLICY(pkey_alloc, ALLOW),
     BPF_SYSCALL_POLICY(pkey_mprotect, TRACE),
diff --git a/runtime/track_memory_map.c b/runtime/track_memory_map.c
index 9fd7a575e..c627f60ea 100644
--- a/runtime/track_memory_map.c
+++ b/runtime/track_memory_map.c
@@ -56,6 +56,11 @@ static bool is_op_permitted(struct memory_map *map, int event,
             map, info->mremap.old_range, info->mremap.pkey))
       return true;
     break;
+  case EVENT_MADVISE:
+    if (memory_map_all_overlapping_regions_have_pkey(
+            map, info->madvise.range, info->madvise.pkey))
+      return true;
+    break;
   case EVENT_MPROTECT: {
     /* allow mprotecting memory that has not been mprotected */
     bool impacts_only_unprotected_memory =
@@ -153,6 +158,11 @@ static bool update_memory_map(struct memory_map *map, int event,
     }
     break;
   }
+  case EVENT_MADVISE:
+    /* madvise does not modify the memory map state we care about here, but can
+    clear memory contents with MADV_DONTNEED */
+    return true;
+    break;
   case EVENT_MPROTECT:
     return memory_map_mprotect_region(map, info->mprotect.range,
                                       info->mprotect.prot);
@@ -282,6 +292,18 @@ static bool interpret_syscall(struct user_regs_struct *regs, unsigned char pkey,
              info->new_range.len);
     break;
   }
+  case EVENT_MADVISE: {
+    struct madvise_info *info = &event_info->madvise;
+    info->range.start = regs->rdi;
+    info->range.len = regs->rsi;
+    info->pkey = pkey;
+
+    int advice = regs->rdx;
+
+    debug_op("compartment %d madvise (%08zx, %zd) with advice=%d\n", info->pkey,
+             info->range.start, info->range.len, advice);
+    break;
+  }
   case EVENT_MPROTECT: {
     struct mprotect_info *info = &event_info->mprotect;
     info->range.start = regs->rdi;