Commit 3015f3d2 authored by Paolo Valente's avatar Paolo Valente Committed by David S. Miller

pkt_sched: enable QFQ to support TSO/GSO

If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.

This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.

The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.

Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.

[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2Signed-off-by: default avatarPaolo Valente <paolo.valente@unimore.it>
Tested-by: default avatarCong Wang <amwang@redhat.com>
Acked-by: default avatarStephen Hemminger <shemminger@vyatta.com>
Acked-by: default avatarStephen Hemminger <shemminger@vyatta.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent cacb6ba0
...@@ -84,18 +84,19 @@ ...@@ -84,18 +84,19 @@
* grp->index is the index of the group; and grp->slot_shift * grp->index is the index of the group; and grp->slot_shift
* is the shift for the corresponding (scaled) sigma_i. * is the shift for the corresponding (scaled) sigma_i.
*/ */
#define QFQ_MAX_INDEX 19 #define QFQ_MAX_INDEX 24
#define QFQ_MAX_WSHIFT 16 #define QFQ_MAX_WSHIFT 12
#define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT) #define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT)
#define QFQ_MAX_WSUM (2*QFQ_MAX_WEIGHT) #define QFQ_MAX_WSUM (16*QFQ_MAX_WEIGHT)
#define FRAC_BITS 30 /* fixed point arithmetic */ #define FRAC_BITS 30 /* fixed point arithmetic */
#define ONE_FP (1UL << FRAC_BITS) #define ONE_FP (1UL << FRAC_BITS)
#define IWSUM (ONE_FP/QFQ_MAX_WSUM) #define IWSUM (ONE_FP/QFQ_MAX_WSUM)
#define QFQ_MTU_SHIFT 11 #define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */
#define QFQ_MIN_SLOT_SHIFT (FRAC_BITS + QFQ_MTU_SHIFT - QFQ_MAX_INDEX) #define QFQ_MIN_SLOT_SHIFT (FRAC_BITS + QFQ_MTU_SHIFT - QFQ_MAX_INDEX)
#define QFQ_MIN_LMAX 256 /* min possible lmax for a class */
/* /*
* Possible group states. These values are used as indexes for the bitmaps * Possible group states. These values are used as indexes for the bitmaps
...@@ -231,6 +232,32 @@ static void qfq_update_class_params(struct qfq_sched *q, struct qfq_class *cl, ...@@ -231,6 +232,32 @@ static void qfq_update_class_params(struct qfq_sched *q, struct qfq_class *cl,
q->wsum += delta_w; q->wsum += delta_w;
} }
static void qfq_update_reactivate_class(struct qfq_sched *q,
struct qfq_class *cl,
u32 inv_w, u32 lmax, int delta_w)
{
bool need_reactivation = false;
int i = qfq_calc_index(inv_w, lmax);
if (&q->groups[i] != cl->grp && cl->qdisc->q.qlen > 0) {
/*
* shift cl->F back, to not charge the
* class for the not-yet-served head
* packet
*/
cl->F = cl->S;
/* remove class from its slot in the old group */
qfq_deactivate_class(q, cl);
need_reactivation = true;
}
qfq_update_class_params(q, cl, lmax, inv_w, delta_w);
if (need_reactivation) /* activate in new group */
qfq_activate_class(q, cl, qdisc_peek_len(cl->qdisc));
}
static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
struct nlattr **tca, unsigned long *arg) struct nlattr **tca, unsigned long *arg)
{ {
...@@ -238,7 +265,7 @@ static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, ...@@ -238,7 +265,7 @@ static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
struct qfq_class *cl = (struct qfq_class *)*arg; struct qfq_class *cl = (struct qfq_class *)*arg;
struct nlattr *tb[TCA_QFQ_MAX + 1]; struct nlattr *tb[TCA_QFQ_MAX + 1];
u32 weight, lmax, inv_w; u32 weight, lmax, inv_w;
int i, err; int err;
int delta_w; int delta_w;
if (tca[TCA_OPTIONS] == NULL) { if (tca[TCA_OPTIONS] == NULL) {
...@@ -270,16 +297,14 @@ static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, ...@@ -270,16 +297,14 @@ static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
if (tb[TCA_QFQ_LMAX]) { if (tb[TCA_QFQ_LMAX]) {
lmax = nla_get_u32(tb[TCA_QFQ_LMAX]); lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
if (!lmax || lmax > (1UL << QFQ_MTU_SHIFT)) { if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
pr_notice("qfq: invalid max length %u\n", lmax); pr_notice("qfq: invalid max length %u\n", lmax);
return -EINVAL; return -EINVAL;
} }
} else } else
lmax = 1UL << QFQ_MTU_SHIFT; lmax = psched_mtu(qdisc_dev(sch));
if (cl != NULL) { if (cl != NULL) {
bool need_reactivation = false;
if (tca[TCA_RATE]) { if (tca[TCA_RATE]) {
err = gen_replace_estimator(&cl->bstats, &cl->rate_est, err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
qdisc_root_sleeping_lock(sch), qdisc_root_sleeping_lock(sch),
...@@ -291,24 +316,8 @@ static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, ...@@ -291,24 +316,8 @@ static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
if (lmax == cl->lmax && inv_w == cl->inv_w) if (lmax == cl->lmax && inv_w == cl->inv_w)
return 0; /* nothing to update */ return 0; /* nothing to update */
i = qfq_calc_index(inv_w, lmax);
sch_tree_lock(sch); sch_tree_lock(sch);
if (&q->groups[i] != cl->grp && cl->qdisc->q.qlen > 0) { qfq_update_reactivate_class(q, cl, inv_w, lmax, delta_w);
/*
* shift cl->F back, to not charge the
* class for the not-yet-served head
* packet
*/
cl->F = cl->S;
/* remove class from its slot in the old group */
qfq_deactivate_class(q, cl);
need_reactivation = true;
}
qfq_update_class_params(q, cl, lmax, inv_w, delta_w);
if (need_reactivation) /* activate in new group */
qfq_activate_class(q, cl, qdisc_peek_len(cl->qdisc));
sch_tree_unlock(sch); sch_tree_unlock(sch);
return 0; return 0;
...@@ -663,15 +672,48 @@ static void qfq_make_eligible(struct qfq_sched *q, u64 old_V) ...@@ -663,15 +672,48 @@ static void qfq_make_eligible(struct qfq_sched *q, u64 old_V)
/* /*
* XXX we should make sure that slot becomes less than 32. * If the weight and lmax (max_pkt_size) of the classes do not change,
* This is guaranteed by the input values. * then QFQ guarantees that the slot index is never higher than
* roundedS is always cl->S rounded on grp->slot_shift bits. * 2 + ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) * (QFQ_MAX_WEIGHT/QFQ_MAX_WSUM).
*
* With the current values of the above constants, the index is
* then guaranteed to never be higher than 2 + 256 * (1 / 16) = 18.
*
* When the weight of a class is increased or the lmax of the class is
* decreased, a new class with smaller slot size may happen to be
* activated. The activation of this class should be properly delayed
* to when the service of the class has finished in the ideal system
* tracked by QFQ. If the activation of the class is not delayed to
* this reference time instant, then this class may be unjustly served
* before other classes waiting for service. This may cause
* (unfrequently) the above bound to the slot index to be violated for
* some of these unlucky classes.
*
* Instead of delaying the activation of the new class, which is quite
* complex, the following inaccurate but simple solution is used: if
* the slot index is higher than QFQ_MAX_SLOTS-2, then the timestamps
* of the class are shifted backward so as to let the slot index
* become equal to QFQ_MAX_SLOTS-2. This threshold is used because, if
* the slot index is above it, then the data structure implementing
* the bucket list either gets immediately corrupted or may get
* corrupted on a possible next packet arrival that causes the start
* time of the group to be shifted backward.
*/ */
static void qfq_slot_insert(struct qfq_group *grp, struct qfq_class *cl, static void qfq_slot_insert(struct qfq_group *grp, struct qfq_class *cl,
u64 roundedS) u64 roundedS)
{ {
u64 slot = (roundedS - grp->S) >> grp->slot_shift; u64 slot = (roundedS - grp->S) >> grp->slot_shift;
unsigned int i = (grp->front + slot) % QFQ_MAX_SLOTS; unsigned int i; /* slot index in the bucket list */
if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
u64 deltaS = roundedS - grp->S -
((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
cl->S -= deltaS;
cl->F -= deltaS;
slot = QFQ_MAX_SLOTS - 2;
}
i = (grp->front + slot) % QFQ_MAX_SLOTS;
hlist_add_head(&cl->next, &grp->slots[i]); hlist_add_head(&cl->next, &grp->slots[i]);
__set_bit(slot, &grp->full_slots); __set_bit(slot, &grp->full_slots);
...@@ -892,6 +934,13 @@ static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch) ...@@ -892,6 +934,13 @@ static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
} }
pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid); pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
if (unlikely(cl->lmax < qdisc_pkt_len(skb))) {
pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
cl->lmax, qdisc_pkt_len(skb), cl->common.classid);
qfq_update_reactivate_class(q, cl, cl->inv_w,
qdisc_pkt_len(skb), 0);
}
err = qdisc_enqueue(skb, cl->qdisc); err = qdisc_enqueue(skb, cl->qdisc);
if (unlikely(err != NET_XMIT_SUCCESS)) { if (unlikely(err != NET_XMIT_SUCCESS)) {
pr_debug("qfq_enqueue: enqueue failed %d\n", err); pr_debug("qfq_enqueue: enqueue failed %d\n", err);
......
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