bpf, events: fix offset in skb copy handler
authorDaniel Borkmann <daniel@iogearbox.net>
Thu, 21 Jul 2016 23:19:42 +0000 (01:19 +0200)
committerDavid S. Miller <davem@davemloft.net>
Mon, 25 Jul 2016 17:34:11 +0000 (10:34 -0700)
This patch fixes the __output_custom() routine we currently use with
bpf_skb_copy(). I missed that when len is larger than the size of the
current handle, we can issue multiple invocations of copy_func, and
__output_custom() advances destination but also source buffer by the
written amount of bytes. When we have __output_custom(), this is actually
wrong since in that case the source buffer points to a non-linear object,
in our case an skb, which the copy_func helper is supposed to walk.
Therefore, since this is non-linear we thus need to pass the offset into
the helper, so that copy_func can use it for extracting the data from
the source object.

Therefore, adjust the callback signatures properly and pass offset
into the skb_header_pointer() invoked from bpf_skb_copy() callback. The
__DEFINE_OUTPUT_COPY_BODY() is adjusted to accommodate for two things:
i) to pass in whether we should advance source buffer or not; this is
a compile-time constant condition, ii) to pass in the offset for
__output_custom(), which we do with help of __VA_ARGS__, so everything
can stay inlined as is currently. Both changes allow for adapting the
__output_* fast-path helpers w/o extra overhead.

Fixes: 555c8a8623a3 ("bpf: avoid stack copy and use skb ctx for event output")
Fixes: 7e3f977edd0b ("perf, events: add non-linear data support for raw records")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/bpf.h
include/linux/perf_event.h
kernel/events/internal.h
net/core/filter.c

index 36da0749205a362a88829b2dc170e245acf7a0c8..11134238417dd4c1054f2cf6c83f5f64b470cb08 100644 (file)
@@ -211,7 +211,7 @@ bool bpf_prog_array_compatible(struct bpf_array *array, const struct bpf_prog *f
 const struct bpf_func_proto *bpf_get_trace_printk_proto(void);
 
 typedef unsigned long (*bpf_ctx_copy_t)(void *dst, const void *src,
-                                       unsigned long len);
+                                       unsigned long off, unsigned long len);
 
 u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
                     void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy);
index e79e6c6fed8984c3c05ca4fa64685b86bfdbe8bf..15e55b7ee0965a137c615097b0919fd8d8fe6b3b 100644 (file)
@@ -70,7 +70,7 @@ struct perf_callchain_entry_ctx {
 };
 
 typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
-                                    unsigned long len);
+                                    unsigned long off, unsigned long len);
 
 struct perf_raw_frag {
        union {
index 2417eb5512cda340fdefe0df469cbf668a03d077..486fd78eb8d5e426dcc88ee6c68e28a10601decf 100644 (file)
@@ -123,18 +123,19 @@ static inline unsigned long perf_aux_size(struct ring_buffer *rb)
        return rb->aux_nr_pages << PAGE_SHIFT;
 }
 
-#define __DEFINE_OUTPUT_COPY_BODY(memcpy_func)                         \
+#define __DEFINE_OUTPUT_COPY_BODY(advance_buf, memcpy_func, ...)       \
 {                                                                      \
        unsigned long size, written;                                    \
                                                                        \
        do {                                                            \
                size    = min(handle->size, len);                       \
-               written = memcpy_func(handle->addr, buf, size);         \
+               written = memcpy_func(__VA_ARGS__);                     \
                written = size - written;                               \
                                                                        \
                len -= written;                                         \
                handle->addr += written;                                \
-               buf += written;                                         \
+               if (advance_buf)                                        \
+                       buf += written;                                 \
                handle->size -= written;                                \
                if (!handle->size) {                                    \
                        struct ring_buffer *rb = handle->rb;            \
@@ -153,12 +154,16 @@ static inline unsigned long perf_aux_size(struct ring_buffer *rb)
 static inline unsigned long                                            \
 func_name(struct perf_output_handle *handle,                           \
          const void *buf, unsigned long len)                           \
-__DEFINE_OUTPUT_COPY_BODY(memcpy_func)
+__DEFINE_OUTPUT_COPY_BODY(true, memcpy_func, handle->addr, buf, size)
 
 static inline unsigned long
 __output_custom(struct perf_output_handle *handle, perf_copy_f copy_func,
                const void *buf, unsigned long len)
-__DEFINE_OUTPUT_COPY_BODY(copy_func)
+{
+       unsigned long orig_len = len;
+       __DEFINE_OUTPUT_COPY_BODY(false, copy_func, handle->addr, buf,
+                                 orig_len - len, size)
+}
 
 static inline unsigned long
 memcpy_common(void *dst, const void *src, unsigned long n)
index 0b521353008df94c4c92b7f7e760928f6fab67e6..5708999f8a7945ec738e2043057817403e1b64ef 100644 (file)
@@ -2026,9 +2026,9 @@ bool bpf_helper_changes_skb_data(void *func)
 }
 
 static unsigned long bpf_skb_copy(void *dst_buff, const void *skb,
-                                 unsigned long len)
+                                 unsigned long off, unsigned long len)
 {
-       void *ptr = skb_header_pointer(skb, 0, len, dst_buff);
+       void *ptr = skb_header_pointer(skb, off, len, dst_buff);
 
        if (unlikely(!ptr))
                return len;