On 3/1/22 10:35, Dumitru Ceara wrote:
On 3/1/22 10:07, Adrian Moreno wrote:


On 2/25/22 18:14, Dumitru Ceara wrote:
For example is parsing the OVN "eth.dst[40] = 1;" action, which
triggered the following warning from UndefinedBehaviorSanitizer:

    lib/meta-flow.c:3210:9: runtime error: member access within
misaligned address 0x000000de4e36 for type 'const union mf_value',
which requires 8 byte alignment
    0x000000de4e36: note: pointer points here
     00 00 00 00 01 00  00 00 00 00 00 00 00 00  70 4e de 00 00 00 00
00  10 51 de 00 00 00 00 00  c0 4f
                 ^
        #0 0x5818bc in mf_format lib/meta-flow.c:3210
        #1 0x5b6047 in format_SET_FIELD lib/ofp-actions.c:3342
        #2 0x5d68ab in ofpact_format lib/ofp-actions.c:9213
        #3 0x5d6ee0 in ofpacts_format lib/ofp-actions.c:9237
        #4 0x410922 in test_parse_actions tests/test-ovn.c:1360

Signed-off-by: Dumitru Ceara <[email protected]>
---
v4: Rebase.
v3: Split out from old patch 07/11.
---
   include/openvswitch/util.h   |    3 +++
   lib/meta-flow.c              |    6 ++++++
   lib/nx-match.c               |    7 +++++++
   ofproto/ofproto-dpif-xlate.c |   13 +++++++++----
   4 files changed, 25 insertions(+), 4 deletions(-)

diff --git a/include/openvswitch/util.h b/include/openvswitch/util.h
index 228b185c3a5f..f45dc505164c 100644
--- a/include/openvswitch/util.h
+++ b/include/openvswitch/util.h
@@ -285,6 +285,9 @@ is_pow2(uintmax_t x)
    * segfault, so it is important to be aware of correct alignment. */
   #define ALIGNED_CAST(TYPE, ATTR) ((TYPE) (void *) (ATTR))
   +#define IS_PTR_ALIGNED(OBJ) \
+    (!(OBJ) || (uintptr_t) (OBJ) % __alignof__(OVS_TYPEOF(OBJ)) == 0)
+
   #ifdef __cplusplus
   }
   #endif
diff --git a/lib/meta-flow.c b/lib/meta-flow.c
index c576ae6202a4..2e64fed5f8f2 100644
--- a/lib/meta-flow.c
+++ b/lib/meta-flow.c
@@ -3172,6 +3172,8 @@ mf_format(const struct mf_field *mf,
             const struct ofputil_port_map *port_map,
             struct ds *s)
   {
+    union mf_value aligned_mask;
+
       if (mask) {
           if (is_all_zeros(mask, mf->n_bytes)) {
               ds_put_cstr(s, "ANY");
@@ -3207,6 +3209,10 @@ mf_format(const struct mf_field *mf,
           break;
         case MFS_ETHERNET:
+        if (!IS_PTR_ALIGNED(mask)) {
+            memcpy(&aligned_mask.mac, mask, sizeof aligned_mask.mac);
+            mask = &aligned_mask;
+        }
           eth_format_masked(value->mac, mask ? &mask->mac : NULL, s);
           break;
   diff --git a/lib/nx-match.c b/lib/nx-match.c
index 440f5f7630c9..6e87deffc0d0 100644
--- a/lib/nx-match.c
+++ b/lib/nx-match.c
@@ -31,6 +31,7 @@
   #include "openvswitch/ofp-match.h"
   #include "openvswitch/ofp-port.h"
   #include "openvswitch/ofpbuf.h"
+#include "openvswitch/util.h"
   #include "openvswitch/vlog.h"
   #include "packets.h"
   #include "openvswitch/shash.h"
@@ -1491,9 +1492,15 @@ nx_put_entry(struct ofpbuf *b, const struct
mf_field *mff,
                enum ofp_version version, const union mf_value *value,
                const union mf_value *mask)
   {
+    union mf_value aligned_mask;
       bool masked;
       int len, offset;
   +    if (!IS_PTR_ALIGNED(mask)) {
+        memcpy(&aligned_mask, mask, mff->n_bytes);
+        mask = &aligned_mask;
+    }
+
       len = mf_field_len(mff, value, mask, &masked);
       offset = mff->n_bytes - len;
   diff --git a/ofproto/ofproto-dpif-xlate.c
b/ofproto/ofproto-dpif-xlate.c
index 5806eacd2e74..c02dcfdc2bf8 100644
--- a/ofproto/ofproto-dpif-xlate.c
+++ b/ofproto/ofproto-dpif-xlate.c
@@ -7124,10 +7124,15 @@ do_xlate_actions(const struct ofpact *ofpacts,
size_t ofpacts_len,
                 /* Set the field only if the packet actually has it. */
               if (mf_are_prereqs_ok(mf, flow, wc)) {
-                mf_mask_field_masked(mf,
ofpact_set_field_mask(set_field), wc);
-                mf_set_flow_value_masked(mf, set_field->value,
-
ofpact_set_field_mask(set_field),
-                                         flow);
+                union mf_value *mask = ofpact_set_field_mask(set_field);
+                union mf_value aligned_mask;
+
+                if (!IS_PTR_ALIGNED(mask)) {
+                    memcpy(&aligned_mask, mask, mf->n_bytes);
+                    mask = &aligned_mask;
+                }
+                mf_mask_field_masked(mf, mask, wc);
+                mf_set_flow_value_masked(mf, set_field->value, mask,
flow);
               } else {
                   xlate_report(ctx, OFT_WARN,
                                "unmet prerequisites for %s, set_field
ignored",


Hi Dumitru,


Hi Adrian,

I'm still trying to understand this patch properly, so don't consider
this a proper review, but I have an initial question.

Does this problem only appear on "eth"?

 From what I can tell, it can only appear on "eth", at least in the
current code.


I can quickly reproduce it with:

./utilities/ovs-ofctl -O OpenFlow15 parse-flow "tcp actions=set_field:0x00ab/0x00ff->tcp_src"

lib/meta-flow.c:1370:37: runtime error: member access within misaligned address 0x000002a247d2 for type 'const union mf_value', which requires 8 byte alignment
0x000002a247d2: note: pointer points here
00 00 00 ab 00 ff 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 31 00 00 00 00 00
              ^

lib/nx-match.c:1501:60: runtime error: member access within misaligned address 0x000002a247d2 for type 'const union mf_value', which requires 8 byte alignment
0x000002a247d2: note: pointer points here
00 00 00 ab 00 ff 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 31 00 00 00 00 00
              ^


Looking at:

/* Use macro to not have to deal with constness. */
#define ofpact_set_field_mask(SF)                               \
     ALIGNED_CAST(union mf_value *,                              \
                  (uint8_t *)(SF)->value + (SF)->field->n_bytes)

It seems that the mask could be misaligned for all fields whose n_bytes
is not 4-byte aligned. If so, shouldn't we deal with this issue in
ofpact_set_field_mask (probably making it a function)?


It might be safer, but do you have a suggestion about how to do this nicely?

I see a few options but I'm not really sure what's best:

a. ofpact_set_field_mask() could always allocate an aligned 'union
mf_value *mask' and return it.  This seems inefficient.

b. ofpact_set_field_mask() could use some per thread static storage and
use that to store an aligned copy of the mask it would normally return
and then return a pointer to that storage.  This seems error-prone, if
the caller retains pointers to the 'mask'.

c. all callers of ofpact_set_field_mask() could pass some extra,
aligned, storage that can be used to store a copy of the mask, if
needed.  This puts the burden on the callers and is also not necessarily
nice and simple to follow.


I agree neither are perfect. I need a bit more thinking.

I'm not familiar with this part of the code but, would it be possible to insert padding between the value and the mask in "ofpact_put_set_field"? I'm not sure if it's possible to add some extra metadata to the "ofpact_set_field" structure so that it can be retrieved by:

>> #define ofpact_set_field_mask(SF)                               \
>>      ALIGNED_CAST(union mf_value *,                             \
>> (uint8_t *)(SF)->value + (SF)->padding + \ (SF)->field->n_bytes)

or maybe just always round up to the nearest 8byte address. Is this possible?

Thanks

Thanks,
Dumitru


Thanks
--
Adrián Moreno

_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to