summaryrefslogtreecommitdiff
path: root/proto_ipv6_no_nxt_hdr.c
blob: 1988422a99ff93d6cf9219143ec2bbd82308e388 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
/*
 * netsniff-ng - the packet sniffing beast
 * Copyright 2012 Markus Amend <markus@netsniff-ng.org>, Deutsche Flugsicherung GmbH
 * Subject to the GPL, version 2.
 *
 * IPv6 No Next Header described in RFC2460
 */

#include <stdio.h>
#include <stdint.h>
#include <netinet/in.h>    /* for ntohs() */

#include "proto.h"
#include "protos.h"
#include "dissector_eth.h"
#include "built_in.h"

static void no_next_header(struct pkt_buff *pkt)
{
	/*
	 * The value 59 in the Next Header field of an IPv6 header or any
	 * extension header indicates that there is nothing following that
	 * header.  If the Payload Length field of the IPv6 header indicates the
	 * presence of octets past the end of a header whose Next Header field
	 * contains 59, those octets must be ignored, and passed on unchanged if
	 * the packet is forwarded.
	 */
	tprintf(" [ No Next Header");
	tprintf(" ]\n");
}

static void no_next_header_less(struct pkt_buff *pkt)
{
	tprintf(" No Next Header");
}

struct protocol ipv6_no_next_header_ops = {
	.key = 0x3B,
	.print_full = no_next_header,
	.print_less = no_next_header_less,
};
s='oid'>6610d0edf6dc7ee97e46ab3a538a565c79d26199 (diff)
dmaengine: pl330: fix double lock
The static bug finder EBA (http://www.iagoabal.eu/eba/) reported the following double-lock bug: Double lock: 1. spin_lock_irqsave(pch->lock, flags) at pl330_free_chan_resources:2236; 2. call to function `pl330_release_channel' immediately after; 3. call to function `dma_pl330_rqcb' in line 1753; 4. spin_lock_irqsave(pch->lock, flags) at dma_pl330_rqcb:1505. I have fixed it as suggested by Marek Szyprowski. First, I have replaced `pch->lock' with `pl330->lock' in functions `pl330_alloc_chan_resources' and `pl330_free_chan_resources'. This avoids the double-lock by acquiring a different lock than `dma_pl330_rqcb'. NOTE that, as a result, `pl330_free_chan_resources' executes `list_splice_tail_init' on `pch->work_list' under lock `pl330->lock', whereas in the rest of the code `pch->work_list' is protected by `pch->lock'. I don't know if this may cause race conditions. Similarly `pch->cyclic' is written by `pl330_alloc_chan_resources' under `pl330->lock' but read by `pl330_tx_submit' under `pch->lock'. Second, I have removed locking from `pl330_request_channel' and `pl330_release_channel' functions. Function `pl330_request_channel' is only called from `pl330_alloc_chan_resources', so the lock is already held. Function `pl330_release_channel' is called from `pl330_free_chan_resources', which already holds the lock, and from `pl330_del'. Function `pl330_del' is called in an error path of `pl330_probe' and at the end of `pl330_remove', but I assume that there cannot be concurrent accesses to the protected data at those points. Signed-off-by: Iago Abal <mail@iagoabal.eu> Reviewed-by: Marek Szyprowski <m.szyprowski@samsung.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
Diffstat (limited to 'net/sched/sch_sfq.c')