Commit 8d174351 authored by Rahul Lakkireddy's avatar Rahul Lakkireddy Committed by David S. Miller

cxgb4: rework TC filter rule insertion across regions

Chelsio NICs have 3 filter regions, in following order of priority:
1. High Priority (HPFILTER) region (Highest Priority).
2. HASH region.
3. Normal FILTER region (Lowest Priority).

Currently, there's a 1-to-1 mapping between the prio value passed
by TC and the filter region index. However, it's possible to have
multiple TC rules with the same prio value. In this case, if a region
is exhausted, no attempt is made to try inserting the rule in the
next available region.

So, rework and remove the 1-to-1 mapping. Instead, dynamically select
the region to insert the filter rule, as long as the new rule's prio
value doesn't conflict with existing rules across all the 3 regions.
Signed-off-by: default avatarRahul Lakkireddy <rahul.lakkireddy@chelsio.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent 357b6cc5
......@@ -438,13 +438,118 @@ int cxgb4_get_filter_counters(struct net_device *dev, unsigned int fidx,
return get_filter_count(adapter, fidx, hitcnt, bytecnt, hash);
}
int cxgb4_get_free_ftid(struct net_device *dev, int family)
static bool cxgb4_filter_prio_in_range(struct tid_info *t, u32 idx, u8 nslots,
u32 prio)
{
struct filter_entry *prev_tab, *next_tab, *prev_fe, *next_fe;
u32 prev_ftid, next_ftid;
/* Only insert the rule if both of the following conditions
* are met:
* 1. The immediate previous rule has priority <= @prio.
* 2. The immediate next rule has priority >= @prio.
*/
/* High Priority (HPFILTER) region always has higher priority
* than normal FILTER region. So, all rules in HPFILTER region
* must have prio value <= rules in normal FILTER region.
*/
if (idx < t->nhpftids) {
/* Don't insert if there's a rule already present at @idx
* in HPFILTER region.
*/
if (test_bit(idx, t->hpftid_bmap))
return false;
next_tab = t->hpftid_tab;
next_ftid = find_next_bit(t->hpftid_bmap, t->nhpftids, idx);
if (next_ftid >= t->nhpftids) {
/* No next entry found in HPFILTER region.
* See if there's any next entry in normal
* FILTER region.
*/
next_ftid = find_first_bit(t->ftid_bmap, t->nftids);
if (next_ftid >= t->nftids)
next_ftid = idx;
else
next_tab = t->ftid_tab;
}
/* Search for the closest previous filter entry in HPFILTER
* region. No need to search in normal FILTER region because
* there can never be any entry in normal FILTER region whose
* prio value is < last entry in HPFILTER region.
*/
prev_ftid = find_last_bit(t->hpftid_bmap, idx);
if (prev_ftid >= idx)
prev_ftid = idx;
prev_tab = t->hpftid_tab;
} else {
idx -= t->nhpftids;
/* Don't insert if there's a rule already present at @idx
* in normal FILTER region.
*/
if (test_bit(idx, t->ftid_bmap))
return false;
prev_tab = t->ftid_tab;
prev_ftid = find_last_bit(t->ftid_bmap, idx);
if (prev_ftid >= idx) {
/* No previous entry found in normal FILTER
* region. See if there's any previous entry
* in HPFILTER region.
*/
prev_ftid = find_last_bit(t->hpftid_bmap, t->nhpftids);
if (prev_ftid >= t->nhpftids)
prev_ftid = idx;
else
prev_tab = t->hpftid_tab;
}
/* Search for the closest next filter entry in normal
* FILTER region. No need to search in HPFILTER region
* because there can never be any entry in HPFILTER
* region whose prio value is > first entry in normal
* FILTER region.
*/
next_ftid = find_next_bit(t->ftid_bmap, t->nftids, idx);
if (next_ftid >= t->nftids)
next_ftid = idx;
next_tab = t->ftid_tab;
}
next_fe = &next_tab[next_ftid];
/* See if the filter entry belongs to an IPv6 rule, which
* occupy 4 slots on T5 and 2 slots on T6. Adjust the
* reference to the previously inserted filter entry
* accordingly.
*/
prev_fe = &prev_tab[prev_ftid & ~(nslots - 1)];
if (!prev_fe->fs.type)
prev_fe = &prev_tab[prev_ftid];
if ((prev_fe->valid && prev_fe->fs.tc_prio > prio) ||
(next_fe->valid && next_fe->fs.tc_prio < prio))
return false;
return true;
}
int cxgb4_get_free_ftid(struct net_device *dev, u8 family, bool hash_en,
u32 tc_prio)
{
struct adapter *adap = netdev2adap(dev);
struct tid_info *t = &adap->tids;
struct filter_entry *tab, *f;
u32 bmap_ftid, max_ftid;
unsigned long *bmap;
bool found = false;
u8 i, n, cnt;
int ftid;
u8 i, cnt, n;
int ftid = 0;
/* IPv4 occupy 1 slot. IPv6 occupy 2 slots on T6 and 4 slots
* on T5.
......@@ -456,34 +561,129 @@ int cxgb4_get_free_ftid(struct net_device *dev, int family)
n += 2;
}
if (n > t->nftids)
return -ENOMEM;
/* Find free filter slots from the end of TCAM. Appropriate
* checks must be done by caller later to ensure the prio
* passed by TC doesn't conflict with prio saved by existing
* rules in the TCAM.
/* There are 3 filter regions available in hardware in
* following order of priority:
*
* 1. High Priority (HPFILTER) region (Highest Priority).
* 2. HASH region.
* 3. Normal FILTER region (Lowest Priority).
*
* Entries in HPFILTER and normal FILTER region have index
* 0 as the highest priority and the rules will be scanned
* in ascending order until either a rule hits or end of
* the region is reached.
*
* All HASH region entries have same priority. The set of
* fields to match in headers are pre-determined. The same
* set of header match fields must be compulsorily specified
* in all the rules wanting to get inserted in HASH region.
* Hence, HASH region is an exact-match region. A HASH is
* generated for a rule based on the values in the
* pre-determined set of header match fields. The generated
* HASH serves as an index into the HASH region. There can
* never be 2 rules having the same HASH. Hardware will
* compute a HASH for every incoming packet based on the
* values in the pre-determined set of header match fields
* and uses it as an index to check if there's a rule
* inserted in the HASH region at the specified index. If
* there's a rule inserted, then it's considered as a filter
* hit. Otherwise, it's a filter miss and normal FILTER region
* is scanned afterwards.
*/
spin_lock_bh(&t->ftid_lock);
ftid = t->nftids - 1;
while (ftid >= n - 1) {
ftid = (tc_prio <= t->nhpftids) ? 0 : t->nhpftids;
max_ftid = t->nftids + t->nhpftids;
while (ftid < max_ftid) {
if (ftid < t->nhpftids) {
/* If the new rule wants to get inserted into
* HPFILTER region, but its prio is greater
* than the rule with the highest prio in HASH
* region, then reject the rule.
*/
if (t->tc_hash_tids_max_prio &&
tc_prio > t->tc_hash_tids_max_prio)
break;
/* If there's not enough slots available
* in HPFILTER region, then move on to
* normal FILTER region immediately.
*/
if (ftid + n > t->nhpftids) {
ftid = t->nhpftids;
continue;
}
bmap = t->hpftid_bmap;
bmap_ftid = ftid;
tab = t->hpftid_tab;
} else if (hash_en) {
/* Ensure priority is >= last rule in HPFILTER
* region.
*/
ftid = find_last_bit(t->hpftid_bmap, t->nhpftids);
if (ftid < t->nhpftids) {
f = &t->hpftid_tab[ftid];
if (f->valid && tc_prio < f->fs.tc_prio)
break;
}
/* Ensure priority is <= first rule in normal
* FILTER region.
*/
ftid = find_first_bit(t->ftid_bmap, t->nftids);
if (ftid < t->nftids) {
f = &t->ftid_tab[ftid];
if (f->valid && tc_prio > f->fs.tc_prio)
break;
}
found = true;
ftid = t->nhpftids;
goto out_unlock;
} else {
/* If the new rule wants to get inserted into
* normal FILTER region, but its prio is less
* than the rule with the highest prio in HASH
* region, then reject the rule.
*/
if (t->tc_hash_tids_max_prio &&
tc_prio < t->tc_hash_tids_max_prio)
break;
if (ftid + n > max_ftid)
break;
bmap = t->ftid_bmap;
bmap_ftid = ftid - t->nhpftids;
tab = t->ftid_tab;
}
cnt = 0;
for (i = 0; i < n; i++) {
if (test_bit(ftid - i, t->ftid_bmap))
if (test_bit(bmap_ftid + i, bmap))
break;
cnt++;
}
if (cnt == n) {
/* Ensure the new rule's prio doesn't conflict
* with existing rules.
*/
if (cxgb4_filter_prio_in_range(t, ftid, n,
tc_prio)) {
ftid &= ~(n - 1);
found = true;
break;
}
}
ftid -= n;
ftid += n;
}
spin_unlock_bh(&t->ftid_lock);
ftid += t->nhpftids;
out_unlock:
spin_unlock_bh(&t->ftid_lock);
return found ? ftid : -ENOMEM;
}
......@@ -555,73 +755,6 @@ static void cxgb4_clear_hpftid(struct tid_info *t, int fidx, int family)
spin_unlock_bh(&t->ftid_lock);
}
bool cxgb4_filter_prio_in_range(struct net_device *dev, u32 idx, u32 prio)
{
struct filter_entry *prev_fe, *next_fe, *tab;
struct adapter *adap = netdev2adap(dev);
u32 prev_ftid, next_ftid, max_tid;
struct tid_info *t = &adap->tids;
unsigned long *bmap;
bool valid = true;
if (idx < t->nhpftids) {
bmap = t->hpftid_bmap;
tab = t->hpftid_tab;
max_tid = t->nhpftids;
} else {
idx -= t->nhpftids;
bmap = t->ftid_bmap;
tab = t->ftid_tab;
max_tid = t->nftids;
}
/* Only insert the rule if both of the following conditions
* are met:
* 1. The immediate previous rule has priority <= @prio.
* 2. The immediate next rule has priority >= @prio.
*/
spin_lock_bh(&t->ftid_lock);
/* Don't insert if there's a rule already present at @idx. */
if (test_bit(idx, bmap)) {
valid = false;
goto out_unlock;
}
next_ftid = find_next_bit(bmap, max_tid, idx);
if (next_ftid >= max_tid)
next_ftid = idx;
next_fe = &tab[next_ftid];
prev_ftid = find_last_bit(bmap, idx);
if (prev_ftid >= idx)
prev_ftid = idx;
/* See if the filter entry belongs to an IPv6 rule, which
* occupy 4 slots on T5 and 2 slots on T6. Adjust the
* reference to the previously inserted filter entry
* accordingly.
*/
if (CHELSIO_CHIP_VERSION(adap->params.chip) < CHELSIO_T6) {
prev_fe = &tab[prev_ftid & ~0x3];
if (!prev_fe->fs.type)
prev_fe = &tab[prev_ftid];
} else {
prev_fe = &tab[prev_ftid & ~0x1];
if (!prev_fe->fs.type)
prev_fe = &tab[prev_ftid];
}
if ((prev_fe->valid && prio < prev_fe->fs.tc_prio) ||
(next_fe->valid && prio > next_fe->fs.tc_prio))
valid = false;
out_unlock:
spin_unlock_bh(&t->ftid_lock);
return valid;
}
/* Delete the filter at a specified index. */
static int del_filter_wr(struct adapter *adapter, int fidx)
{
......
......@@ -53,5 +53,4 @@ void clear_all_filters(struct adapter *adapter);
void init_hash_filter(struct adapter *adap);
bool is_filter_exact_match(struct adapter *adap,
struct ch_filter_specification *fs);
bool cxgb4_filter_prio_in_range(struct net_device *dev, u32 idx, u32 prio);
#endif /* __CXGB4_FILTER_H */
......@@ -635,6 +635,64 @@ int cxgb4_validate_flow_actions(struct net_device *dev,
return 0;
}
static void cxgb4_tc_flower_hash_prio_add(struct adapter *adap, u32 tc_prio)
{
spin_lock_bh(&adap->tids.ftid_lock);
if (adap->tids.tc_hash_tids_max_prio < tc_prio)
adap->tids.tc_hash_tids_max_prio = tc_prio;
spin_unlock_bh(&adap->tids.ftid_lock);
}
static void cxgb4_tc_flower_hash_prio_del(struct adapter *adap, u32 tc_prio)
{
struct tid_info *t = &adap->tids;
struct ch_tc_flower_entry *fe;
struct rhashtable_iter iter;
u32 found = 0;
spin_lock_bh(&t->ftid_lock);
/* Bail if the current rule is not the one with the max
* prio.
*/
if (t->tc_hash_tids_max_prio != tc_prio)
goto out_unlock;
/* Search for the next rule having the same or next lower
* max prio.
*/
rhashtable_walk_enter(&adap->flower_tbl, &iter);
do {
rhashtable_walk_start(&iter);
fe = rhashtable_walk_next(&iter);
while (!IS_ERR_OR_NULL(fe)) {
if (fe->fs.hash &&
fe->fs.tc_prio <= t->tc_hash_tids_max_prio) {
t->tc_hash_tids_max_prio = fe->fs.tc_prio;
found++;
/* Bail if we found another rule
* having the same prio as the
* current max one.
*/
if (fe->fs.tc_prio == tc_prio)
break;
}
fe = rhashtable_walk_next(&iter);
}
rhashtable_walk_stop(&iter);
} while (fe == ERR_PTR(-EAGAIN));
rhashtable_walk_exit(&iter);
if (!found)
t->tc_hash_tids_max_prio = 0;
out_unlock:
spin_unlock_bh(&t->ftid_lock);
}
int cxgb4_tc_flower_replace(struct net_device *dev,
struct flow_cls_offload *cls)
{
......@@ -644,6 +702,7 @@ int cxgb4_tc_flower_replace(struct net_device *dev,
struct ch_tc_flower_entry *ch_flower;
struct ch_filter_specification *fs;
struct filter_ctx ctx;
u8 inet_family;
int fidx, ret;
if (cxgb4_validate_flow_actions(dev, &rule->action, extack))
......@@ -664,39 +723,32 @@ int cxgb4_tc_flower_replace(struct net_device *dev,
cxgb4_process_flow_actions(dev, &rule->action, fs);
fs->hash = is_filter_exact_match(adap, fs);
if (fs->hash) {
fidx = 0;
} else {
u8 inet_family;
inet_family = fs->type ? PF_INET6 : PF_INET;
/* Note that TC uses prio 0 to indicate stack to
* generate automatic prio and hence doesn't pass prio
* 0 to driver. However, the hardware TCAM index
* starts from 0. Hence, the -1 here.
/* Get a free filter entry TID, where we can insert this new
* rule. Only insert rule if its prio doesn't conflict with
* existing rules.
*/
if (cls->common.prio <= (adap->tids.nftids +
adap->tids.nhpftids)) {
fidx = cls->common.prio - 1;
if (fidx < adap->tids.nhpftids)
fs->prio = 1;
} else {
fidx = cxgb4_get_free_ftid(dev, inet_family);
}
/* Only insert FLOWER rule if its priority doesn't
* conflict with existing rules in the LETCAM.
*/
if (fidx < 0 ||
!cxgb4_filter_prio_in_range(dev, fidx, cls->common.prio)) {
fidx = cxgb4_get_free_ftid(dev, inet_family, fs->hash,
cls->common.prio);
if (fidx < 0) {
NL_SET_ERR_MSG_MOD(extack,
"No free LETCAM index available");
ret = -ENOMEM;
goto free_entry;
}
if (fidx < adap->tids.nhpftids) {
fs->prio = 1;
fs->hash = 0;
}
/* If the rule can be inserted into HASH region, then ignore
* the index to normal FILTER region.
*/
if (fs->hash)
fidx = 0;
fs->tc_prio = cls->common.prio;
fs->tc_cookie = cls->cookie;
......@@ -727,6 +779,9 @@ int cxgb4_tc_flower_replace(struct net_device *dev,
if (ret)
goto del_filter;
if (fs->hash)
cxgb4_tc_flower_hash_prio_add(adap, cls->common.prio);
return 0;
del_filter:
......@@ -742,12 +797,17 @@ int cxgb4_tc_flower_destroy(struct net_device *dev,
{
struct adapter *adap = netdev2adap(dev);
struct ch_tc_flower_entry *ch_flower;
u32 tc_prio;
bool hash;
int ret;
ch_flower = ch_flower_lookup(adap, cls->cookie);
if (!ch_flower)
return -ENOENT;
hash = ch_flower->fs.hash;
tc_prio = ch_flower->fs.tc_prio;
ret = cxgb4_del_filter(dev, ch_flower->filter_id, &ch_flower->fs);
if (ret)
goto err;
......@@ -760,6 +820,9 @@ int cxgb4_tc_flower_destroy(struct net_device *dev,
}
kfree_rcu(ch_flower, rcu);
if (hash)
cxgb4_tc_flower_hash_prio_del(adap, tc_prio);
err:
return ret;
}
......
......@@ -198,22 +198,14 @@ static int cxgb4_matchall_alloc_filter(struct net_device *dev,
struct ch_filter_specification *fs;
int ret, fidx;
/* Note that TC uses prio 0 to indicate stack to generate
* automatic prio and hence doesn't pass prio 0 to driver.
* However, the hardware TCAM index starts from 0. Hence, the
* -1 here. 1 slot is enough to create a wildcard matchall
* VIID rule.
/* Get a free filter entry TID, where we can insert this new
* rule. Only insert rule if its prio doesn't conflict with
* existing rules.
*
* 1 slot is enough to create a wildcard matchall VIID rule.
*/
if (cls->common.prio <= (adap->tids.nftids + adap->tids.nhpftids))
fidx = cls->common.prio - 1;
else
fidx = cxgb4_get_free_ftid(dev, PF_INET);
/* Only insert MATCHALL rule if its priority doesn't conflict
* with existing rules in the LETCAM.
*/
if (fidx < 0 ||
!cxgb4_filter_prio_in_range(dev, fidx, cls->common.prio)) {
fidx = cxgb4_get_free_ftid(dev, PF_INET, false, cls->common.prio);
if (fidx < 0) {
NL_SET_ERR_MSG_MOD(extack,
"No free LETCAM index available");
return -ENOMEM;
......
......@@ -155,9 +155,10 @@ int cxgb4_config_knode(struct net_device *dev, struct tc_cls_u32_offload *cls)
struct ch_filter_specification fs;
struct cxgb4_tc_u32_table *t;
struct cxgb4_link *link;
unsigned int filter_id;
u32 uhtid, link_uhtid;
bool is_ipv6 = false;
u8 inet_family;
int filter_id;
int ret;
if (!can_tc_u32_offload(dev))
......@@ -166,18 +167,15 @@ int cxgb4_config_knode(struct net_device *dev, struct tc_cls_u32_offload *cls)
if (protocol != htons(ETH_P_IP) && protocol != htons(ETH_P_IPV6))
return -EOPNOTSUPP;
/* Note that TC uses prio 0 to indicate stack to generate
* automatic prio and hence doesn't pass prio 0 to driver.
* However, the hardware TCAM index starts from 0. Hence, the
* -1 here.
*/
filter_id = TC_U32_NODE(cls->knode.handle) - 1;
inet_family = (protocol == htons(ETH_P_IPV6)) ? PF_INET6 : PF_INET;
/* Only insert U32 rule if its priority doesn't conflict with
* existing rules in the LETCAM.
/* Get a free filter entry TID, where we can insert this new
* rule. Only insert rule if its prio doesn't conflict with
* existing rules.
*/
if (filter_id >= adapter->tids.nftids + adapter->tids.nhpftids ||
!cxgb4_filter_prio_in_range(dev, filter_id, cls->common.prio)) {
filter_id = cxgb4_get_free_ftid(dev, inet_family, false,
TC_U32_NODE(cls->knode.handle));
if (filter_id < 0) {
NL_SET_ERR_MSG_MOD(extack,
"No free LETCAM index available");
return -ENOMEM;
......@@ -358,23 +356,65 @@ int cxgb4_delete_knode(struct net_device *dev, struct tc_cls_u32_offload *cls)
struct cxgb4_link *link = NULL;
struct cxgb4_tc_u32_table *t;
struct filter_entry *f;
bool found = false;
u32 handle, uhtid;
u8 nslots;
int ret;
if (!can_tc_u32_offload(dev))
return -EOPNOTSUPP;
/* Fetch the location to delete the filter. */
filter_id = TC_U32_NODE(cls->knode.handle) - 1;
if (filter_id >= adapter->tids.nftids + adapter->tids.nhpftids)
return -ERANGE;
max_tids = adapter->tids.nhpftids + adapter->tids.nftids;
spin_lock_bh(&adapter->tids.ftid_lock);
filter_id = 0;
while (filter_id < max_tids) {
if (filter_id < adapter->tids.nhpftids) {
i = filter_id;
f = &adapter->tids.hpftid_tab[i];
if (f->valid && f->fs.tc_cookie == cls->knode.handle) {
found = true;
break;
}
if (filter_id < adapter->tids.nhpftids)
f = &adapter->tids.hpftid_tab[filter_id];
else
f = &adapter->tids.ftid_tab[filter_id - adapter->tids.nhpftids];
i = find_next_bit(adapter->tids.hpftid_bmap,
adapter->tids.nhpftids, i + 1);
if (i >= adapter->tids.nhpftids) {
filter_id = adapter->tids.nhpftids;
continue;
}
if (cls->knode.handle != f->fs.tc_cookie)
filter_id = i;
} else {
i = filter_id - adapter->tids.nhpftids;
f = &adapter->tids.ftid_tab[i];
if (f->valid && f->fs.tc_cookie == cls->knode.handle) {
found = true;
break;
}
i = find_next_bit(adapter->tids.ftid_bmap,
adapter->tids.nftids, i + 1);
if (i >= adapter->tids.nftids)
break;
filter_id = i + adapter->tids.nhpftids;
}
nslots = 0;
if (f->fs.type) {
nslots++;
if (CHELSIO_CHIP_VERSION(adapter->params.chip) <
CHELSIO_T6)
nslots += 2;
}
filter_id += nslots;
}
spin_unlock_bh(&adapter->tids.ftid_lock);
if (!found)
return -ERANGE;
t = adapter->tc_u32;
......@@ -407,7 +447,6 @@ int cxgb4_delete_knode(struct net_device *dev, struct tc_cls_u32_offload *cls)
/* If a link is being deleted, then delete all filters
* associated with the link.
*/
max_tids = adapter->tids.nftids;
for (i = 0; i < t->size; i++) {
link = &t->table[i];
......
......@@ -149,6 +149,8 @@ struct tid_info {
atomic_t conns_in_use;
/* lock for setting/clearing filter bitmap */
spinlock_t ftid_lock;
unsigned int tc_hash_tids_max_prio;
};
static inline void *lookup_tid(const struct tid_info *t, unsigned int tid)
......@@ -263,7 +265,8 @@ struct filter_ctx {
struct ch_filter_specification;
int cxgb4_get_free_ftid(struct net_device *dev, int family);
int cxgb4_get_free_ftid(struct net_device *dev, u8 family, bool hash_en,
u32 tc_prio);
int __cxgb4_set_filter(struct net_device *dev, int filter_id,
struct ch_filter_specification *fs,
struct filter_ctx *ctx);
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment