net/tap: fix eBPF RSS map key handling

This commit addresses a case of RSS and non RSS flows mixture.
In the corrupted code, if a non-RSS flow is destroyed, it has an eBPF map
index 0 (initialized upon flow creation). This might unintentionally remove
RSS entry 0 from eBPF map.
To fix this issue, add an offset to the real index during a KEY_CMD_GET
operation, and subtract this offset during a KEY_CMD_RELEASE operation, in
order to restore the real index.
Thus, if a non RSS flow is falsely trying to release map entry 0 - The
offset subtraction will calculate the real map index as an
out-of-range value, and the release operation will be silently ignored.

Fixes: 036d721a82 ("net/tap: implement RSS using eBPF")

Signed-off-by: Ophir Munk <ophirmu@mellanox.com>
Acked-by: Pascal Mazon <pascal.mazon@6wind.com>
This commit is contained in:
Ophir Munk 2018-01-30 16:00:29 +00:00 committed by Ferruh Yigit
parent da8841a713
commit 1911c5edc6

View File

@ -1797,6 +1797,7 @@ tap_flow_implicit_flush(struct pmd_internals *pmd, struct rte_flow_error *error)
}
#define MAX_RSS_KEYS 256
#define KEY_IDX_OFFSET (3 * MAX_RSS_KEYS)
#define SEC_NAME_CLS_Q "cls_q"
const char *sec_name[SEC_MAX] = {
@ -1953,38 +1954,62 @@ static int rss_enable(struct pmd_internals *pmd,
static int bpf_rss_key(enum bpf_rss_key_e cmd, __u32 *key_idx)
{
__u32 i;
int err = -1;
int err = 0;
static __u32 num_used_keys;
static __u32 rss_keys[MAX_RSS_KEYS] = {KEY_STAT_UNSPEC};
static __u32 rss_keys_initialized;
switch (cmd) {
case KEY_CMD_GET:
if (!rss_keys_initialized)
if (!rss_keys_initialized) {
err = -1;
break;
}
if (num_used_keys == RTE_DIM(rss_keys))
if (num_used_keys == RTE_DIM(rss_keys)) {
err = -1;
break;
}
*key_idx = num_used_keys % RTE_DIM(rss_keys);
while (rss_keys[*key_idx] == KEY_STAT_USED)
*key_idx = (*key_idx + 1) % RTE_DIM(rss_keys);
rss_keys[*key_idx] = KEY_STAT_USED;
/*
* Add an offset to key_idx in order to handle a case of
* RSS and non RSS flows mixture.
* If a non RSS flow is destroyed it has an eBPF map
* index 0 (initialized on flow creation) and might
* unintentionally remove RSS entry 0 from eBPF map.
* To avoid this issue, add an offset to the real index
* during a KEY_CMD_GET operation and subtract this offset
* during a KEY_CMD_RELEASE operation in order to restore
* the real index.
*/
*key_idx += KEY_IDX_OFFSET;
num_used_keys++;
err = 0;
break;
case KEY_CMD_RELEASE:
if (!rss_keys_initialized) {
err = 0;
if (!rss_keys_initialized)
break;
}
if (rss_keys[*key_idx] == KEY_STAT_USED) {
rss_keys[*key_idx] = KEY_STAT_AVAILABLE;
/*
* Subtract offest to restore real key index
* If a non RSS flow is falsely trying to release map
* entry 0 - the offset subtraction will calculate the real
* map index as an out-of-range value and the release operation
* will be silently ignored.
*/
__u32 key = *key_idx - KEY_IDX_OFFSET;
if (key >= RTE_DIM(rss_keys))
break;
if (rss_keys[key] == KEY_STAT_USED) {
rss_keys[key] = KEY_STAT_AVAILABLE;
num_used_keys--;
err = 0;
}
break;
@ -1994,7 +2019,6 @@ static int bpf_rss_key(enum bpf_rss_key_e cmd, __u32 *key_idx)
rss_keys_initialized = 1;
num_used_keys = 0;
err = 0;
break;
case KEY_CMD_DEINIT:
@ -2003,7 +2027,6 @@ static int bpf_rss_key(enum bpf_rss_key_e cmd, __u32 *key_idx)
rss_keys_initialized = 0;
num_used_keys = 0;
err = 0;
break;
default: