Disallow reading object data from Parcels with non-object reads
The check added to each non-object reads adds an overhead. If the
objects (binders and file descriptors) were written to the Parcel in
sequential order then check adds a small O(1) overhead to each read,
plus an O(N) overhead to the first read (to verify the N objects were
added in order).
If the objects were written out of order (as in by jumping around the Parcel
with setDataPosition and writing Binder, DON'T DO THIS!!) (writing non
objects out of order is fine), the first read is forced to sort the objects
in the internal bookkeeping. Based on the assumption non sequential writes
are infrequent and overall Parcels are probably mostly sorted, insertion
sort was used. Worst case sorts will add an O(N^2) overhead to the first
non object read from the Parcel.
Test: run cts -m CtsOsTestCases -t android.os.cts.ParcelTest
Bug: 29833520
Change-Id: I82de8eb5f5eb56f869542d5358e96884c24301b2
(cherry picked from commit c517681c66a1a387be657e0cf06da8d19659dd14)
diff --git a/libs/binder/Parcel.cpp b/libs/binder/Parcel.cpp
index e22179b..44357c3 100644
--- a/libs/binder/Parcel.cpp
+++ b/libs/binder/Parcel.cpp
@@ -433,6 +433,7 @@
mDataPos = pos;
mNextObjectHint = 0;
+ mObjectsSorted = false;
}
status_t Parcel::setDataCapacity(size_t size)
@@ -1469,6 +1470,59 @@
LOG_ALWAYS_FATAL("Parcel::remove() not yet implemented!");
}
+status_t Parcel::validateReadData(size_t upperBound) const
+{
+ // Don't allow non-object reads on object data
+ if (mObjectsSorted || mObjectsSize <= 1) {
+data_sorted:
+ // Expect to check only against the next object
+ if (mNextObjectHint < mObjectsSize && upperBound > mObjects[mNextObjectHint]) {
+ // For some reason the current read position is greater than the next object
+ // hint. Iterate until we find the right object
+ size_t nextObject = mNextObjectHint;
+ do {
+ if (mDataPos < mObjects[nextObject] + sizeof(flat_binder_object)) {
+ // Requested info overlaps with an object
+ ALOGE("Attempt to read from protected data in Parcel %p", this);
+ return PERMISSION_DENIED;
+ }
+ nextObject++;
+ } while (nextObject < mObjectsSize && upperBound > mObjects[nextObject]);
+ mNextObjectHint = nextObject;
+ }
+ return NO_ERROR;
+ }
+ // Quickly determine if mObjects is sorted.
+ binder_size_t* currObj = mObjects + mObjectsSize - 1;
+ binder_size_t* prevObj = currObj;
+ while (currObj > mObjects) {
+ prevObj--;
+ if(*prevObj > *currObj) {
+ goto data_unsorted;
+ }
+ currObj--;
+ }
+ mObjectsSorted = true;
+ goto data_sorted;
+
+data_unsorted:
+ // Insertion Sort mObjects
+ // Great for mostly sorted lists. If randomly sorted or reverse ordered mObjects become common,
+ // switch to std::sort(mObjects, mObjects + mObjectsSize);
+ for (binder_size_t* iter0 = mObjects + 1; iter0 < mObjects + mObjectsSize; iter0++) {
+ binder_size_t temp = *iter0;
+ binder_size_t* iter1 = iter0 - 1;
+ while (iter1 >= mObjects && *iter1 > temp) {
+ *(iter1 + 1) = *iter1;
+ iter1--;
+ }
+ *(iter1 + 1) = temp;
+ }
+ mNextObjectHint = 0;
+ mObjectsSorted = true;
+ goto data_sorted;
+}
+
status_t Parcel::read(void* outData, size_t len) const
{
if (len > INT32_MAX) {
@@ -1479,6 +1533,10 @@
if ((mDataPos+pad_size(len)) >= mDataPos && (mDataPos+pad_size(len)) <= mDataSize
&& len <= pad_size(len)) {
+ if (mObjectsSize > 0) {
+ status_t err = validateReadData(mDataPos + pad_size(len));
+ if(err != NO_ERROR) return err;
+ }
memcpy(outData, mData+mDataPos, len);
mDataPos += pad_size(len);
ALOGV("read Setting data pos of %p to %zu", this, mDataPos);
@@ -1497,6 +1555,11 @@
if ((mDataPos+pad_size(len)) >= mDataPos && (mDataPos+pad_size(len)) <= mDataSize
&& len <= pad_size(len)) {
+ if (mObjectsSize > 0) {
+ status_t err = validateReadData(mDataPos + pad_size(len));
+ if(err != NO_ERROR) return NULL;
+ }
+
const void* data = mData+mDataPos;
mDataPos += pad_size(len);
ALOGV("readInplace Setting data pos of %p to %zu", this, mDataPos);
@@ -1510,6 +1573,11 @@
COMPILE_TIME_ASSERT_FUNCTION_SCOPE(PAD_SIZE_UNSAFE(sizeof(T)) == sizeof(T));
if ((mDataPos+sizeof(T)) <= mDataSize) {
+ if (mObjectsSize > 0) {
+ status_t err = validateReadData(mDataPos + sizeof(T));
+ if(err != NO_ERROR) return err;
+ }
+
const void* data = mData+mDataPos;
mDataPos += sizeof(T);
*pArg = *reinterpret_cast<const T*>(data);
@@ -2366,6 +2434,7 @@
mObjects = const_cast<binder_size_t*>(objects);
mObjectsSize = mObjectsCapacity = objectsCount;
mNextObjectHint = 0;
+ mObjectsSorted = false;
mOwner = relFunc;
mOwnerCookie = relCookie;
for (size_t i = 0; i < mObjectsSize; i++) {
@@ -2524,6 +2593,7 @@
mObjects = NULL;
mObjectsSize = mObjectsCapacity = 0;
mNextObjectHint = 0;
+ mObjectsSorted = false;
mHasFds = false;
mFdsKnown = true;
mAllowFds = true;
@@ -2610,6 +2680,7 @@
mDataCapacity = desired;
mObjectsSize = mObjectsCapacity = objectsSize;
mNextObjectHint = 0;
+ mObjectsSorted = false;
} else if (mData) {
if (objectsSize < mObjectsSize) {
@@ -2631,6 +2702,7 @@
}
mObjectsSize = objectsSize;
mNextObjectHint = 0;
+ mObjectsSorted = false;
}
// We own the data, so we can just do a realloc().
@@ -2703,6 +2775,7 @@
mObjectsSize = 0;
mObjectsCapacity = 0;
mNextObjectHint = 0;
+ mObjectsSorted = false;
mHasFds = false;
mFdsKnown = true;
mAllowFds = true;
diff --git a/libs/binder/include/binder/Parcel.h b/libs/binder/include/binder/Parcel.h
index 5d36526..dede78f 100644
--- a/libs/binder/include/binder/Parcel.h
+++ b/libs/binder/include/binder/Parcel.h
@@ -417,6 +417,7 @@
void freeDataNoInit();
void initState();
void scanForFds() const;
+ status_t validateReadData(size_t len) const;
template<class T>
status_t readAligned(T *pArg) const;
@@ -463,6 +464,7 @@
size_t mObjectsSize;
size_t mObjectsCapacity;
mutable size_t mNextObjectHint;
+ mutable bool mObjectsSorted;
mutable bool mFdsKnown;
mutable bool mHasFds;
