| Message ID | 1534547305-25140-4-git-send-email-tushar.n.dave@oracle.com |
|---|---|
| State | RFC, archived |
| Delegated to: | BPF Maintainers |
| Headers | show |
| Series | eBPF and struct scatterlist | expand |
On 08/17/2018 04:08 PM, Tushar Dave wrote: > Like sockmap (sk_msg), socksg also deals with struct scatterlist > therefore socksg programs can use existing bpf helper bpf_msg_pull_data > to access packet data contained in struct scatterlist. While doing some > prelimnary testing, there are couple of issues found with > bpf_msg_pull_data that are fixed in this patch. > > Also, there cannot be more than MAX_SKB_FRAGS entries in sg_data > therefore any checks for sg entry more than MAX_SKB_FRAGS in > bpf_msg_pull_data() is removed. In sockmap the scatterlist is used as a ring so the MAX_SKB_FRAGS check is needed to keep searching through the ring when sg_start is non-zero. > > Besides that, I also ran into issues while put_page() is invoked. > e.g. > [ 450.568723] BUG: Bad page state in process swapper/10 pfn:2021540 > [ 450.575632] page:ffffea0080855000 count:0 mapcount:0 > mapping:ffff88103d006840 index:0xffff882021540000 compound_mapcount: 0 > [ 450.588069] flags: 0x6fffff80008100(slab|head) > [ 450.593033] raw: 006fffff80008100 dead000000000100 dead000000000200 > ffff88103d006840 > [ 450.601683] raw: ffff882021540000 0000000080080007 00000000ffffffff > 0000000000000000 > [ 450.610337] page dumped because: PAGE_FLAGS_CHECK_AT_FREE flag(s) set > [ 450.617530] bad because of flags: 0x100(slab) > > To avoid above issue, currently put_page() is disabled in this patch > temporarily. I am working on alternatives so that page allocated via > slab (in this case) can be freed without any issue.> > Signed-off-by: Tushar Dave <tushar.n.dave@oracle.com> > Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com> > --- > net/core/filter.c | 61 +++++++++++++++++++++++++++++-------------------------- > 1 file changed, 32 insertions(+), 29 deletions(-) > > diff --git a/net/core/filter.c b/net/core/filter.c > index e427c8e..cc52baa 100644 > --- a/net/core/filter.c > +++ b/net/core/filter.c > @@ -2316,7 +2316,7 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) > BPF_CALL_4(bpf_msg_pull_data, > struct sk_msg_buff *, msg, u32, start, u32, end, u64, flags) > { > - unsigned int len = 0, offset = 0, copy = 0; > + unsigned int len = 0, offset = 0, copy = 0, off = 0; > struct scatterlist *sg = msg->sg_data; > int first_sg, last_sg, i, shift; > unsigned char *p, *to, *from; > @@ -2330,22 +2330,28 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) > i = msg->sg_start; > do { > len = sg[i].length; > - offset += len; > if (start < offset + len) > break; > + offset += len; This looks like a generic fix unrelated to this series. Can you send that as a bugfix? > i++; > - if (i == MAX_SKB_FRAGS) > - i = 0; > - } while (i != msg->sg_end); > + } while (i <= msg->sg_end); > As noted above the MAX_SKB_FRAGS check is needed because sg_start can be non-zero and sg_end < st_start. In these cases we need to search the entries at the start of the array (being used as a ring). > + /* return error if start is out of range */ > if (unlikely(start >= offset + len)) > return -EINVAL; > > - if (!msg->sg_copy[i] && bytes <= len) > - goto out; > + /* return error if i is last entry in sglist and end is out of range */ > + if (msg->sg_copy[i] && end > offset + len) > + return -EINVAL> > first_sg = i; > > + /* if i is not last entry in sg list and end (i.e start + bytes) is > + * within this sg[i] then goto out and calculate data and data_end > + */ > + if (!msg->sg_copy[i] && end <= offset + len) > + goto out; > +> /* At this point we need to linearize multiple scatterlist > * elements or a single shared page. Either way we need to > * copy into a linear buffer exclusively owned by BPF. Then > @@ -2359,11 +2365,14 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) > do { > copy += sg[i].length; > i++; > - if (i == MAX_SKB_FRAGS) > - i = 0; same as above, need to keep. > - if (bytes < copy) > + if (end < copy) > break; > - } while (i != msg->sg_end); > + } while (i <= msg->sg_end); > + > + /* return error if i is last entry in sglist and end is out of range */ > + if (i > msg->sg_end && end > offset + copy) > + return -EINVAL; > + > last_sg = i; > > if (unlikely(copy < end - start)) > @@ -2373,23 +2382,25 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) > if (unlikely(!page)) > return -ENOMEM; > p = page_address(page); > - offset = 0; > > i = first_sg; > do { > from = sg_virt(&sg[i]); > len = sg[i].length; > - to = p + offset; > + to = p + off; Not really sure if the change from offset->off is needed. Looks like it just makes a bigger diff. > > memcpy(to, from, len); > - offset += len; > + off += len; > sg[i].length = 0; > - put_page(sg_page(&sg[i])); > + /* if original page is allocated via slab then put_page > + * causes error BUG: Bad page state in process. So temporarily > + * disabled put_page. > + * Todo: fix it > + */ > + //put_page(sg_page(&sg[i])); > > i++; > - if (i == MAX_SKB_FRAGS) > - i = 0; > - } while (i != last_sg); > + } while (i < last_sg); > > sg[first_sg].length = copy; > sg_set_page(&sg[first_sg], page, copy, 0); > @@ -2406,12 +2417,8 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) > do { > int move_from; > > - if (i + shift >= MAX_SKB_FRAGS) > - move_from = i + shift - MAX_SKB_FRAGS; > - else > - move_from = i + shift; > - Need to keep same as above. > - if (move_from == msg->sg_end) > + move_from = i + shift;> + if (move_from > msg->sg_end) > break; > > sg[i] = sg[move_from]; > @@ -2420,14 +2427,10 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) > sg[move_from].offset = 0; > > i++; > - if (i == MAX_SKB_FRAGS) > - i = 0; > } while (1); > msg->sg_end -= shift; > - if (msg->sg_end < 0) > - msg->sg_end += MAX_SKB_FRAGS; > out: > - msg->data = sg_virt(&sg[i]) + start - offset; > + msg->data = sg_virt(&sg[first_sg]) + start - offset; > msg->data_end = msg->data + bytes; > > return 0; > Thanks, John
On 08/24/2018 06:02 PM, John Fastabend wrote: > On 08/17/2018 04:08 PM, Tushar Dave wrote: >> Like sockmap (sk_msg), socksg also deals with struct scatterlist >> therefore socksg programs can use existing bpf helper bpf_msg_pull_data >> to access packet data contained in struct scatterlist. While doing some >> prelimnary testing, there are couple of issues found with >> bpf_msg_pull_data that are fixed in this patch. >> >> Also, there cannot be more than MAX_SKB_FRAGS entries in sg_data >> therefore any checks for sg entry more than MAX_SKB_FRAGS in >> bpf_msg_pull_data() is removed. > > In sockmap the scatterlist is used as a ring so the MAX_SKB_FRAGS > check is needed to keep searching through the ring when sg_start > is non-zero. Okay. > >> >> Besides that, I also ran into issues while put_page() is invoked. >> e.g. >> [ 450.568723] BUG: Bad page state in process swapper/10 pfn:2021540 >> [ 450.575632] page:ffffea0080855000 count:0 mapcount:0 >> mapping:ffff88103d006840 index:0xffff882021540000 compound_mapcount: 0 >> [ 450.588069] flags: 0x6fffff80008100(slab|head) >> [ 450.593033] raw: 006fffff80008100 dead000000000100 dead000000000200 >> ffff88103d006840 >> [ 450.601683] raw: ffff882021540000 0000000080080007 00000000ffffffff >> 0000000000000000 >> [ 450.610337] page dumped because: PAGE_FLAGS_CHECK_AT_FREE flag(s) set >> [ 450.617530] bad because of flags: 0x100(slab) >> >> To avoid above issue, currently put_page() is disabled in this patch >> temporarily. I am working on alternatives so that page allocated via >> slab (in this case) can be freed without any issue.> >> Signed-off-by: Tushar Dave <tushar.n.dave@oracle.com> >> Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com> >> --- >> net/core/filter.c | 61 +++++++++++++++++++++++++++++-------------------------- >> 1 file changed, 32 insertions(+), 29 deletions(-) >> >> diff --git a/net/core/filter.c b/net/core/filter.c >> index e427c8e..cc52baa 100644 >> --- a/net/core/filter.c >> +++ b/net/core/filter.c >> @@ -2316,7 +2316,7 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) >> BPF_CALL_4(bpf_msg_pull_data, >> struct sk_msg_buff *, msg, u32, start, u32, end, u64, flags) >> { >> - unsigned int len = 0, offset = 0, copy = 0; >> + unsigned int len = 0, offset = 0, copy = 0, off = 0; >> struct scatterlist *sg = msg->sg_data; >> int first_sg, last_sg, i, shift; >> unsigned char *p, *to, *from; >> @@ -2330,22 +2330,28 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) >> i = msg->sg_start; >> do { >> len = sg[i].length; >> - offset += len; >> if (start < offset + len) >> break; >> + offset += len; > > This looks like a generic fix unrelated to this series. > Can you send that as a bugfix? Okay. > >> i++; >> - if (i == MAX_SKB_FRAGS) >> - i = 0; >> - } while (i != msg->sg_end); >> + } while (i <= msg->sg_end); >> > > As noted above the MAX_SKB_FRAGS check is needed because > sg_start can be non-zero and sg_end < st_start. In these > cases we need to search the entries at the start of the > array (being used as a ring). Yup! > >> + /* return error if start is out of range */ >> if (unlikely(start >= offset + len)) >> return -EINVAL; >> >> - if (!msg->sg_copy[i] && bytes <= len) >> - goto out; >> + /* return error if i is last entry in sglist and end is out of range */ >> + if (msg->sg_copy[i] && end > offset + len) >> + return -EINVAL> >> first_sg = i; >> >> + /* if i is not last entry in sg list and end (i.e start + bytes) is >> + * within this sg[i] then goto out and calculate data and data_end >> + */ >> + if (!msg->sg_copy[i] && end <= offset + len) >> + goto out; >> +> /* At this point we need to linearize multiple scatterlist >> * elements or a single shared page. Either way we need to >> * copy into a linear buffer exclusively owned by BPF. Then >> @@ -2359,11 +2365,14 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) >> do { >> copy += sg[i].length; >> i++; >> - if (i == MAX_SKB_FRAGS) >> - i = 0; > > same as above, need to keep. Yup! > >> - if (bytes < copy) >> + if (end < copy) >> break; >> - } while (i != msg->sg_end); >> + } while (i <= msg->sg_end); >> + >> + /* return error if i is last entry in sglist and end is out of range */ >> + if (i > msg->sg_end && end > offset + copy) >> + return -EINVAL; >> + >> last_sg = i; >> >> if (unlikely(copy < end - start)) >> @@ -2373,23 +2382,25 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) >> if (unlikely(!page)) >> return -ENOMEM; >> p = page_address(page); >> - offset = 0; >> >> i = first_sg; >> do { >> from = sg_virt(&sg[i]); >> len = sg[i].length; >> - to = p + offset; >> + to = p + off; > > Not really sure if the change from offset->off is needed. Looks > like it just makes a bigger diff. We need both offset and off because they both are used for different calculations! 'offset' is used to calculate the 'msg->data' i.e. msg->data = sg_virt(&sg[first_sg]) + start - offset" 'off' , on the other hand, is used for when we linearize sg. > >> >> memcpy(to, from, len); >> - offset += len; >> + off += len; >> sg[i].length = 0; >> - put_page(sg_page(&sg[i])); >> + /* if original page is allocated via slab then put_page >> + * causes error BUG: Bad page state in process. So temporarily >> + * disabled put_page. >> + * Todo: fix it >> + */ >> + //put_page(sg_page(&sg[i])); As I said in the commit message that put_page() causes error "BUG: Bad page state in process ..." when used for RDS. Any clue? Have you seen something like this with sockmap? >> >> i++; >> - if (i == MAX_SKB_FRAGS) >> - i = 0; >> - } while (i != last_sg); >> + } while (i < last_sg); >> >> sg[first_sg].length = copy; >> sg_set_page(&sg[first_sg], page, copy, 0); >> @@ -2406,12 +2417,8 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) >> do { >> int move_from; >> >> - if (i + shift >= MAX_SKB_FRAGS) >> - move_from = i + shift - MAX_SKB_FRAGS; >> - else >> - move_from = i + shift; >> - > > Need to keep same as above. yup! > >> - if (move_from == msg->sg_end) >> + move_from = i + shift;> + if (move_from > msg->sg_end) >> break; >> >> sg[i] = sg[move_from]; >> @@ -2420,14 +2427,10 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) >> sg[move_from].offset = 0; >> >> i++; >> - if (i == MAX_SKB_FRAGS) >> - i = 0; >> } while (1); >> msg->sg_end -= shift; >> - if (msg->sg_end < 0) >> - msg->sg_end += MAX_SKB_FRAGS; >> out: >> - msg->data = sg_virt(&sg[i]) + start - offset; >> + msg->data = sg_virt(&sg[first_sg]) + start - offset; >> msg->data_end = msg->data + bytes; >> >> return 0; >> > > Thanks, > John > Thanks. -Tushar
diff --git a/net/core/filter.c b/net/core/filter.c index e427c8e..cc52baa 100644 --- a/net/core/filter.c +++ b/net/core/filter.c @@ -2316,7 +2316,7 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) BPF_CALL_4(bpf_msg_pull_data, struct sk_msg_buff *, msg, u32, start, u32, end, u64, flags) { - unsigned int len = 0, offset = 0, copy = 0; + unsigned int len = 0, offset = 0, copy = 0, off = 0; struct scatterlist *sg = msg->sg_data; int first_sg, last_sg, i, shift; unsigned char *p, *to, *from; @@ -2330,22 +2330,28 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) i = msg->sg_start; do { len = sg[i].length; - offset += len; if (start < offset + len) break; + offset += len; i++; - if (i == MAX_SKB_FRAGS) - i = 0; - } while (i != msg->sg_end); + } while (i <= msg->sg_end); + /* return error if start is out of range */ if (unlikely(start >= offset + len)) return -EINVAL; - if (!msg->sg_copy[i] && bytes <= len) - goto out; + /* return error if i is last entry in sglist and end is out of range */ + if (msg->sg_copy[i] && end > offset + len) + return -EINVAL; first_sg = i; + /* if i is not last entry in sg list and end (i.e start + bytes) is + * within this sg[i] then goto out and calculate data and data_end + */ + if (!msg->sg_copy[i] && end <= offset + len) + goto out; + /* At this point we need to linearize multiple scatterlist * elements or a single shared page. Either way we need to * copy into a linear buffer exclusively owned by BPF. Then @@ -2359,11 +2365,14 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) do { copy += sg[i].length; i++; - if (i == MAX_SKB_FRAGS) - i = 0; - if (bytes < copy) + if (end < copy) break; - } while (i != msg->sg_end); + } while (i <= msg->sg_end); + + /* return error if i is last entry in sglist and end is out of range */ + if (i > msg->sg_end && end > offset + copy) + return -EINVAL; + last_sg = i; if (unlikely(copy < end - start)) @@ -2373,23 +2382,25 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) if (unlikely(!page)) return -ENOMEM; p = page_address(page); - offset = 0; i = first_sg; do { from = sg_virt(&sg[i]); len = sg[i].length; - to = p + offset; + to = p + off; memcpy(to, from, len); - offset += len; + off += len; sg[i].length = 0; - put_page(sg_page(&sg[i])); + /* if original page is allocated via slab then put_page + * causes error BUG: Bad page state in process. So temporarily + * disabled put_page. + * Todo: fix it + */ + //put_page(sg_page(&sg[i])); i++; - if (i == MAX_SKB_FRAGS) - i = 0; - } while (i != last_sg); + } while (i < last_sg); sg[first_sg].length = copy; sg_set_page(&sg[first_sg], page, copy, 0); @@ -2406,12 +2417,8 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) do { int move_from; - if (i + shift >= MAX_SKB_FRAGS) - move_from = i + shift - MAX_SKB_FRAGS; - else - move_from = i + shift; - - if (move_from == msg->sg_end) + move_from = i + shift; + if (move_from > msg->sg_end) break; sg[i] = sg[move_from]; @@ -2420,14 +2427,10 @@ struct sock *do_msg_redirect_map(struct sk_msg_buff *msg) sg[move_from].offset = 0; i++; - if (i == MAX_SKB_FRAGS) - i = 0; } while (1); msg->sg_end -= shift; - if (msg->sg_end < 0) - msg->sg_end += MAX_SKB_FRAGS; out: - msg->data = sg_virt(&sg[i]) + start - offset; + msg->data = sg_virt(&sg[first_sg]) + start - offset; msg->data_end = msg->data + bytes; return 0;