Re: [Qemu-devel] [PATCH v3 4/5] fw_cfg: add generic non-DMA read method

[Gabriel L. Somlo]

On Tue, Nov 03, 2015 at 11:53:55AM +0100, Laszlo Ersek wrote:
> Thank you for splitting out this patch; it makes it easier to review.
> However,
> 
> On 11/03/15 01:35, Gabriel L. Somlo wrote:
> > Introduce fw_cfg_data_read(), a generic read method which works
> > on all access widths (1 through 8 bytes, inclusive), and can be
> > used during both IOPort and MMIO read accesses.
> > 
> > To maintain legibility, only fw_cfg_data_mem_read() (the MMIO
> > data read method) is replaced by this patch. The new method
> > essentially unwinds the fw_cfg_data_mem_read() + fw_cfg_read()
> > combo, but without unnecessarily repeating all the validity
> > checks performed by the latter on each byte being read.
> 
> this unwinding caused a bug to creep in.
> 
> Namely, we have to identify the set of data that remains constant
> between *all* "size" calls that fw_cfg_data_mem_read() makes to
> fw_cfg_read(), and hoist / eliminate the checks on those *only*.
> 
> Specifically,
> 
> > This patch also modifies the trace_fw_cfg_read prototype to
> > accept a 64-bit value argument, allowing it to work properly
> > with the new read method, but also remain backward compatible
> > with existing call sites.
> > 
> > Cc: Laszlo Ersek <address@hidden>
> > Cc: Gerd Hoffmann <address@hidden>
> > Cc: Marc Marí <address@hidden>
> > Signed-off-by: Gabriel Somlo <address@hidden>
> > ---
> >  hw/nvram/fw_cfg.c | 33 +++++++++++++++++++--------------
> >  trace-events      |  2 +-
> >  2 files changed, 20 insertions(+), 15 deletions(-)
> > 
> > diff --git a/hw/nvram/fw_cfg.c b/hw/nvram/fw_cfg.c
> > index c2d3a0a..8aa980c 100644
> > --- a/hw/nvram/fw_cfg.c
> > +++ b/hw/nvram/fw_cfg.c
> > @@ -274,6 +274,24 @@ static int fw_cfg_select(FWCfgState *s, uint16_t key)
> >      return ret;
> >  }
> >  
> > +static uint64_t fw_cfg_data_read(void *opaque, hwaddr addr, unsigned size)
> > +{
> > +    FWCfgState *s = opaque;
> 
> This is good.
> 
> > +    int arch = !!(s->cur_entry & FW_CFG_ARCH_LOCAL);
> 
> Okay too.
> 
> > +    FWCfgEntry *e = &s->entries[arch][s->cur_entry & FW_CFG_ENTRY_MASK];
> 
> (1) Side point: the conversion here is faithful to the original code in
> fw_cfg_read(), but even in the original code, the expression uses
> "s->cur_entry" as a (masked) subscript *before* comparing it against
> FW_CFG_INVALID. I don't think that's right.
> 
> The same issue is present in fw_cfg_dma_transfer(). Care to write a
> patch (before the restructuring) that fixes both?
> 
> Note, I am aware that the expression in both of the above mentioned
> functions only calculates the *address* of the nonexistent element
> belonging to (FW_CFG_INVALID & FW_CFG_ENTRY_MASK) == 0x3FFF:
> 
>   e = &s->entries[arch][s->cur_entry & FW_CFG_ENTRY_MASK];
> 
> But it doesn't matter; it's undefined behavior just the same. Instead,
> *both* locations should say:
> 
>  e = (s->cur_entry == FW_CFG_INVALID) ? NULL :
>      &s->entries[arch][s->cur_entry & FW_CFG_ENTRY_MASK];
> 
> (I share the blame for not noticing this earlier -- I too reviewed
> fw_cfg_dma_transfer().)
> 
> NULL is a valid pointer to *evaluate* (not to dereference), whereas the
> current address-of expression is not valid even for evaluation. Also, in
> practice, dereferencing NULL would give us a nice (as in, non-garbage)
> SIGSEGV.

Done.

> 
> Anyway, back to the topic at hand:
> 
> > +    uint64_t value = 0;
> > +
> > +    assert(size <= sizeof(value));
> > +    if (s->cur_entry != FW_CFG_INVALID && e->data) {
> 
> Right, good conversion. (Side note: this does protect against
> *dereferencing* "e", but it's already too late, as far as undefined
> behavior is concerned.)
> 
> > +        while (size-- && s->cur_offset < e->len) {
> > +            value = (value << 8) | e->data[s->cur_offset++];
> > +        }
> 
> (2) So, this is the bug. The pre-conversion code would keep shifting
> "value" to the left until "size" was reached, regardless of the
> underlying blob size, and just leave the least significant bytes zeroed
> if the item ended too early. Whereas this loop *stops shifting* when the
> blob ends.

D'OH!!! That should teach me to pay more attention -- thanks for
catching it!

> Since the wide data register (which is big-endian) implements a
> substring-preserving transfer (on top of QEMU's integer preserving
> device r/w infrastructure), this change breaks the case when the
> firmware reads, say, 8 bytes from the register in a single access, when
> only 3 are left in the blob, and then uses only the three *lowest
> address* bytes from the uint64_t value read. Although no known firmware
> does this at the moment, it would be valid, and the above hunk would
> break it.
> 
> Hence please
> 
> (2a) either append the missing "cumulative" shift after the loop:
> 
>     while (size && s->cur_offset < e->len) {
>         --size;
>         value = (value << 8) | e->data[s->cur_offset++];
>     }
>     value <<= 8 * size;

I went with 2a. Also added a comment to make things painfully obvious
to any potential future archaeologists:

+static uint64_t fw_cfg_data_read(void *opaque, hwaddr addr, unsigned size)
+{
+    FWCfgState *s = opaque;
+    int arch = !!(s->cur_entry & FW_CFG_ARCH_LOCAL);
+    FWCfgEntry *e = (s->cur_entry == FW_CFG_INVALID) ? NULL :
+                    &s->entries[arch][s->cur_entry & FW_CFG_ENTRY_MASK];
+    uint64_t value = 0;
+
+    assert(size <= sizeof(value));
+    if (s->cur_entry != FW_CFG_INVALID && e->data) {
+        /* The least significant 'size' bytes of the return value are
+         * expected to contain a string preserving portion of the item
+         * data, padded with zeros to the right in case we run out early.
+         */
+        while (size && s->cur_offset < e->len) {
+            value = (value << 8) | e->data[s->cur_offset++];
+            size--;
+        }
+        /* If size is still not zero, we *did* run out early, so finish
+         * left-shifting to add the appropriate number of padding zeros
+         * on the right.
+         */
+        value <<= 8 * size;
+    }
+
+    trace_fw_cfg_read(s, value);
+    return value;
+}

Version 4 should be out by the end of today.

Thanks again,
--Gabriel

> 
> (2b) or move the offset check from the loop's controlling expression
> into the value composition:
> 
>         while (size--) {
>             value = (value << 8) | (s->cur_offset < e->len ?
>                                     e->data[s->cur_offset++] :
>                                     0);
>         }
> 
> The rest looks good.
> 
> Thanks
> Laszlo
> 
> > +    }
> > +
> > +    trace_fw_cfg_read(s, value);
> > +    return value;
> > +}
> > +
> >  static uint8_t fw_cfg_read(FWCfgState *s)
> >  {
> >      int arch = !!(s->cur_entry & FW_CFG_ARCH_LOCAL);
> > @@ -290,19 +308,6 @@ static uint8_t fw_cfg_read(FWCfgState *s)
> >      return ret;
> >  }
> >  
> > -static uint64_t fw_cfg_data_mem_read(void *opaque, hwaddr addr,
> > -                                     unsigned size)
> > -{
> > -    FWCfgState *s = opaque;
> > -    uint64_t value = 0;
> > -    unsigned i;
> > -
> > -    for (i = 0; i < size; ++i) {
> > -        value = (value << 8) | fw_cfg_read(s);
> > -    }
> > -    return value;
> > -}
> > -
> >  static void fw_cfg_data_mem_write(void *opaque, hwaddr addr,
> >                                    uint64_t value, unsigned size)
> >  {
> > @@ -483,7 +488,7 @@ static const MemoryRegionOps fw_cfg_ctl_mem_ops = {
> >  };
> >  
> >  static const MemoryRegionOps fw_cfg_data_mem_ops = {
> > -    .read = fw_cfg_data_mem_read,
> > +    .read = fw_cfg_data_read,
> >      .write = fw_cfg_data_mem_write,
> >      .endianness = DEVICE_BIG_ENDIAN,
> >      .valid = {
> > diff --git a/trace-events b/trace-events
> > index 72136b9..5073040 100644
> > --- a/trace-events
> > +++ b/trace-events
> > @@ -196,7 +196,7 @@ ecc_diag_mem_readb(uint64_t addr, uint32_t ret) "Read 
> > diagnostic %"PRId64"= %02x
> >  
> >  # hw/nvram/fw_cfg.c
> >  fw_cfg_select(void *s, uint16_t key, int ret) "%p key %d = %d"
> > -fw_cfg_read(void *s, uint8_t ret) "%p = %d"
> > +fw_cfg_read(void *s, uint64_t ret) "%p = %"PRIx64
> >  fw_cfg_add_file(void *s, int index, char *name, size_t len) "%p #%d: %s 
> > (%zd bytes)"
> >  
> >  # hw/block/hd-geometry.c
> > 
>