[PATCH 04/12] fuzz: Add DMA support to the generic-fuzzer

[Alexander Bulekov <alxndr@bu.edu>]

When a virtual-device tries to access some buffer in memory over DMA, we
add call-backs into the fuzzer(next commit). The fuzzer checks verifies
that the DMA request maps to a physical RAM address and fills the memory
with fuzzer-provided data. The patterns that we use to fill this memory
are specified using add_dma_pattern and clear_dma_patterns operations.

Signed-off-by: Alexander Bulekov <alxndr@bu.edu>
---
 tests/qtest/fuzz/general_fuzz.c | 177 ++++++++++++++++++++++++++++++++
 1 file changed, 177 insertions(+)

diff --git a/tests/qtest/fuzz/general_fuzz.c b/tests/qtest/fuzz/general_fuzz.c
index e715b77d59..4b6967c5d2 100644
--- a/tests/qtest/fuzz/general_fuzz.c
+++ b/tests/qtest/fuzz/general_fuzz.c
@@ -27,6 +27,7 @@
 #include "tests/qtest/libqos/pci.h"
 #include "tests/qtest/libqos/pci-pc.h"
 #include "hw/pci/pci.h"
+#include "hw/boards.h"
 
 /*
  * CMD_SEP is a random 32-bit value used to separate "commands" in the fuzz
@@ -34,6 +35,7 @@
  */
 #define CMD_SEP "\x84\x05\x5C\x5E"
 #define DEFAULT_TIMEOUT_US 100000
+#define MAX_DMA_FILL_SIZE 0x10000
 
 #define PCI_HOST_BRIDGE_CFG 0xcf8
 #define PCI_HOST_BRIDGE_DATA 0xcfc
@@ -44,6 +46,24 @@ typedef struct {
 } address_range;
 
 static useconds_t timeout = 100000;
+/*
+ * A pattern used to populate a DMA region or perform a memwrite. This is
+ * useful for e.g. populating tables of unique addresses.
+ * Example {.index = 1; .stride = 2; .len = 3; .data = "\x00\x01\x02"}
+ * Renders as: 00 01 02   00 03 03   00 05 03   00 07 03 ...
+ */
+typedef struct {
+    uint8_t index;      /* Index of a byte to increment by stride */
+    uint8_t stride;     /* Increment each index'th byte by this amount */
+    size_t len;
+    const uint8_t *data;
+} pattern;
+
+/* Avoid filling the same DMA region between MMIO/PIO commands ? */
+static bool avoid_double_fetches;
+
+static QTestState *qts_global; /* Need a global for the DMA callback */
+
 /*
  * List of memory regions that are children of QOM objects specified by the
  * user for fuzzing.
@@ -51,6 +71,122 @@ static useconds_t timeout = 100000;
 static GPtrArray *fuzzable_memoryregions;
 static GPtrArray *fuzzable_pci_devices;
 
+/*
+ * List of dma regions populated since the last fuzzing command. Used to ensure
+ * that we only write to each DMA address once, to avoid race conditions when
+ * building reproducers.
+ */
+static GArray *dma_regions;
+
+static GArray *dma_patterns;
+int dma_pattern_index;
+
+/*
+ * Allocate a block of memory and populate it with a pattern.
+ */
+static void *pattern_alloc(pattern p, size_t len)
+{
+    int i;
+    uint8_t *buf = g_malloc(len);
+    uint8_t sum = 0;
+
+    for (i = 0; i < len; ++i) {
+        buf[i] = p.data[i % p.len];
+        if ((i % p.len) == p.index) {
+            buf[i] += sum;
+            sum += p.stride;
+        }
+    }
+    return buf;
+}
+
+/*
+ * Call-back for functions that perform DMA reads from guest memory. Confirm
+ * that the region has not already been populated since the last loop in
+ * general_fuzz(), avoiding potential race-conditions, which we don't have
+ * a good way for reproducing right now.
+ */
+void fuzz_dma_read_cb(size_t addr, size_t len, MemoryRegion *mr, bool is_write)
+{
+    /* Are we in the general-fuzzer or are we using another fuzz-target? */
+    if (!qts_global) {
+        return;
+    }
+
+    /*
+     * If the device is trying to read from a ROM, exit early. We do not want
+     * to fuzz devices using data that we have no control over.
+     */
+    if (mr->readonly) {
+        _Exit(0);
+    }
+
+    /*
+     * Return immediately if:
+     * - We have no DMA patterns defined
+     * - The length of the DMA read request is zero
+     * - The DMA read is hitting an MR other than the machine's main RAM
+     * - The DMA request is not a read (what happens for a address_space_map
+     *   with is_write=True? Can the device use the same pointer to do reads?)
+     * - The DMA request hits past the bounds of our RAM
+     */
+    if (dma_patterns->len == 0
+        || len == 0
+        || mr != MACHINE(qdev_get_machine())->ram
+        || is_write
+        || addr > current_machine->ram_size) {
+        return;
+    }
+
+    /*
+     * If we overlap with any existing dma_regions, split the range and only
+     * populate the non-overlapping parts.
+     */
+    for (int i = 0; i < dma_regions->len && !avoid_double_fetches; ++i) {
+        address_range region = g_array_index(dma_regions, address_range, i);
+        if (addr < region.addr + region.len && addr + len > region.addr) {
+            if (addr < region.addr) {
+                fuzz_dma_read_cb(addr, region.addr - addr, mr, is_write);
+            }
+            if (addr + len > region.addr + region.len) {
+                fuzz_dma_read_cb(region.addr + region.len,
+                        addr + len - (region.addr + region.len), mr, is_write);
+            }
+            return;
+        }
+    }
+
+    /* Cap the length of the DMA access to something reasonable */
+    len = MIN(len, MAX_DMA_FILL_SIZE);
+
+    address_range ar = {addr, len};
+    g_array_append_val(dma_regions, ar);
+    pattern p = g_array_index(dma_patterns, pattern, dma_pattern_index);
+    void *buf = pattern_alloc(p, ar.len);
+    if (getenv("QTEST_LOG")) {
+        /*
+         * With QTEST_LOG, use a normal, slow QTest memwrite. Prefix the log
+         * that will be written by qtest.c with a DMA tag, so we can reorder
+         * the resulting QTest trace so the DMA fills precede the last PIO/MMIO
+         * command.
+         */
+        fprintf(stderr, "[DMA] ");
+        fflush(stderr);
+        qtest_memwrite(qts_global, ar.addr, buf, ar.len);
+    } else {
+       /*
+        * Populate the region using address_space_write_rom to avoid writing to
+        * any IO MemoryRegions
+        */
+        address_space_write_rom(first_cpu->as, ar.addr, MEMTXATTRS_UNSPECIFIED,
+                buf, ar.len);
+    }
+    free(buf);
+
+    /* Increment the index of the pattern for the next DMA access */
+    dma_pattern_index = (dma_pattern_index + 1) % dma_patterns->len;
+}
+
 /*
  * Here we want to convert a fuzzer-provided [io-region-index, offset] to
  * a physical address. To do this, we iterate over all of the matched
@@ -334,6 +470,35 @@ static void op_pci_write(QTestState *s, const unsigned char * data, size_t len)
     }
 }
 
+static void op_add_dma_pattern(QTestState *s,
+                               const unsigned char *data, size_t len)
+{
+    struct {
+        /*
+         * index and stride can be used to increment the index-th byte of the
+         * pattern by the value stride, for each loop of the pattern.
+         */
+        uint8_t index;
+        uint8_t stride;
+    } a;
+
+    if (len < sizeof(a) + 1) {
+        return;
+    }
+    memcpy(&a, data, sizeof(a));
+    pattern p = {a.index, a.stride, len - sizeof(a), data + sizeof(a)};
+    p.index = a.index % p.len;
+    g_array_append_val(dma_patterns, p);
+    return;
+}
+
+static void op_clear_dma_patterns(QTestState *s,
+                                  const unsigned char *data, size_t len)
+{
+    g_array_set_size(dma_patterns, 0);
+    dma_pattern_index = 0;
+}
+
 static void op_clock_step(QTestState *s, const unsigned char *data, size_t len)
 {
     qtest_clock_step_next(s);
@@ -380,6 +545,8 @@ static void general_fuzz(QTestState *s, const unsigned char *Data, size_t Size)
         op_write,
         op_pci_read,
         op_pci_write,
+        op_add_dma_pattern,
+        op_clear_dma_patterns,
         op_clock_step,
     };
     const unsigned char *cmd = Data;
@@ -433,6 +600,9 @@ static void usage(void)
     printf("QEMU_FUZZ_ARGS= the command line arguments passed to qemu\n");
     printf("QEMU_FUZZ_OBJECTS= "
             "a space separated list of QOM type names for objects to fuzz\n");
+    printf("Optionally: QEMU_AVOID_DOUBLE_FETCH= "
+            "Try to avoid racy DMA double fetch bugs? %d by default\n",
+            avoid_double_fetches);
     printf("Optionally: QEMU_FUZZ_TIMEOUT= Specify a custom timeout (us). "
             "0 to disable. %d by default\n", timeout);
     exit(0);
@@ -502,9 +672,16 @@ static void general_pre_fuzz(QTestState *s)
     if (!getenv("QEMU_FUZZ_OBJECTS")) {
         usage();
     }
+    if (getenv("QEMU_AVOID_DOUBLE_FETCH")) {
+        avoid_double_fetches = 1;
+    }
     if (getenv("QEMU_FUZZ_TIMEOUT")) {
         timeout = g_ascii_strtoll(getenv("QEMU_FUZZ_TIMEOUT"), NULL, 0);
     }
+    qts_global = s;
+
+    dma_regions = g_array_new(false, false, sizeof(address_range));
+    dma_patterns = g_array_new(false, false, sizeof(pattern));
 
     fuzzable_memoryregions = g_ptr_array_new();
     fuzzable_pci_devices = g_ptr_array_new();
-- 
2.27.0