Re: [PATCH v2 04/15] fuzz: Add DMA support to the generic-fuzzer

[Alexander Bulekov]

On 200903 0943, Darren Kenny wrote:
> On Wednesday, 2020-08-19 at 02:10:59 -04, Alexander Bulekov wrote:
> > 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 | 178 ++++++++++++++++++++++++++++++++
> >  1 file changed, 178 insertions(+)
> >
> > diff --git a/tests/qtest/fuzz/general_fuzz.c 
> > b/tests/qtest/fuzz/general_fuzz.c
> > index 17b572a439..36d41acea0 100644
> > --- a/tests/qtest/fuzz/general_fuzz.c
> > +++ b/tests/qtest/fuzz/general_fuzz.c
> > @@ -25,6 +25,8 @@
> >  #include "exec/address-spaces.h"
> >  #include "hw/qdev-core.h"
> >  #include "hw/pci/pci.h"
> > +#include "hw/boards.h"
> > +#include "exec/memory-internal.h"
> >  
> >  /*
> >   * SEPARATOR is used to separate "operations" in the fuzz input
> > @@ -38,12 +40,16 @@ enum cmds{
> >      OP_WRITE,
> >      OP_PCI_READ,
> >      OP_PCI_WRITE,
> > +    OP_ADD_DMA_PATTERN,
> > +    OP_CLEAR_DMA_PATTERNS,
> >      OP_CLOCK_STEP,
> >  };
> >  
> >  #define DEFAULT_TIMEOUT_US 100000
> >  #define USEC_IN_SEC 100000000
> >  
> > +#define MAX_DMA_FILL_SIZE 0x10000
> > +
> >  #define PCI_HOST_BRIDGE_CFG 0xcf8
> >  #define PCI_HOST_BRIDGE_DATA 0xcfc
> >  
> > @@ -53,6 +59,24 @@ 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 ...
> 
> TYPO: I think this wrong, and that the last digit should be 02 not 03 in
>       each group.
>

Ah thanks for catching that.

> > + */
> > +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.
> > @@ -60,6 +84,116 @@ 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;
> > +static int dma_pattern_index;
> > +
> > +void fuzz_dma_read_cb(size_t addr, size_t len, MemoryRegion *mr, bool 
> > is_write);
> > +
> > +/*
> > + * 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;
> > +    }
> > +
> > +    /*
> > +     * 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 && !(mr->ops == 
> > &unassigned_mem_ops))
> > +        || 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);
> 
> NIT: Can be slightly more expensive to declare a variable on each
> iteration, but also tends to be cleaner not to do this.

Ok, I'll fix that.

> > +        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")) {
> 
> NIT: It might be cleaner to put any testing of env vars in to
> the code in general_fuzz() where most others are being tested, and
> instead set a static global boolean which can be used here instead.
> Depending on how many times this is called, it may also be slightly
> faster since getenv() has to search an array of strings, etc. to get the
> value.

True. I think the env-variables, in-general, need to be
handled/documented in some consistent location.

> > +        /*
> > +         * 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);
> 
> NIT: For consistency, this probably should be a g_free(), since the memory
> was allocated using g_malloc().
> 

Oops - I'll fix that.

> > +
> > +    /* 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
> > @@ -350,6 +484,35 @@ 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);
> > @@ -396,6 +559,8 @@ static void general_fuzz(QTestState *s, const unsigned 
> > char *Data, size_t Size)
> >          [OP_WRITE]              = op_write,
> >          [OP_PCI_READ]           = op_pci_read,
> >          [OP_PCI_WRITE]          = op_pci_write,
> > +        [OP_ADD_DMA_PATTERN]    = op_add_dma_pattern,
> > +        [OP_CLEAR_DMA_PATTERNS] = op_clear_dma_patterns,
> >          [OP_CLOCK_STEP]         = op_clock_step,
> >      };
> >      const unsigned char *cmd = Data;
> > @@ -425,6 +590,8 @@ static void general_fuzz(QTestState *s, const unsigned 
> > char *Data, size_t Size)
> >              setitimer(ITIMER_VIRTUAL, &timer, NULL);
> >          }
> >  
> > +        op_clear_dma_patterns(s, NULL, 0);
> > +
> >          while (cmd && Size) {
> >              /* Get the length until the next command or end of input */
> >              nextcmd = memmem(cmd, Size, SEPARATOR, strlen(SEPARATOR));
> > @@ -441,6 +608,7 @@ static void general_fuzz(QTestState *s, const unsigned 
> > char *Data, size_t Size)
> >              /* Advance to the next command */
> >              cmd = nextcmd ? nextcmd + sizeof(SEPARATOR) - 1 : nextcmd;
> >              Size = Size - (cmd_len + sizeof(SEPARATOR) - 1);
> > +            g_array_set_size(dma_regions, 0);
> >          }
> >          _Exit(0);
> >      } else {
> > @@ -455,6 +623,9 @@ 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);
> > @@ -522,9 +693,16 @@ 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();
> 
> Since mostly nits and typos:
> 
> Reviewed-by: Darren Kenny <darren.kenny@oracle.com>
> 
> Thanks,
> 
> Darren.