| 1 | From 9d02daf754238adac48fa075ee79e7edd3d79ed3 Mon Sep 17 00:00:00 2001 |
| 2 | From: David Woodhouse <dwmw2@infradead.org> |
| 3 | Date: Sun, 8 Apr 2012 09:55:43 +0000 |
| 4 | Subject: [PATCH] pppoatm: Fix excessive queue bloat |
| 5 | |
| 6 | We discovered that PPPoATM has an excessively deep transmit queue. A |
| 7 | queue the size of the default socket send buffer (wmem_default) is |
| 8 | maintained between the PPP generic core and the ATM device. |
| 9 | |
| 10 | Fix it to queue a maximum of *two* packets. The one the ATM device is |
| 11 | currently working on, and one more for the ATM driver to process |
| 12 | immediately in its TX done interrupt handler. The PPP core is designed |
| 13 | to feed packets to the channel with minimal latency, so that really |
| 14 | ought to be enough to keep the ATM device busy. |
| 15 | |
| 16 | While we're at it, fix the fact that we were triggering the wakeup |
| 17 | tasklet on *every* pppoatm_pop() call. The comment saying "this is |
| 18 | inefficient, but doing it right is too hard" turns out to be overly |
| 19 | pessimistic... I think :) |
| 20 | |
| 21 | On machines like the Traverse Geos, with a slow Geode CPU and two |
| 22 | high-speed ADSL2+ interfaces, there were reports of extremely high CPU |
| 23 | usage which could partly be attributed to the extra wakeups. |
| 24 | |
| 25 | (The wakeup handling could actually be made a whole lot easier if we |
| 26 | stop checking sk->sk_sndbuf altogether. Given that we now only queue |
| 27 | *two* packets ever, one wonders what the point is. As it is, you could |
| 28 | already deadlock the thing by setting the sk_sndbuf to a value lower |
| 29 | than the MTU of the device, and it'd just block for ever.) |
| 30 | |
| 31 | Signed-off-by: David Woodhouse <David.Woodhouse@intel.com> |
| 32 | Signed-off-by: David S. Miller <davem@davemloft.net> |
| 33 | --- |
| 34 | net/atm/pppoatm.c | 95 +++++++++++++++++++++++++++++++++++++++++++++++----- |
| 35 | 1 files changed, 85 insertions(+), 10 deletions(-) |
| 36 | |
| 37 | --- a/net/atm/pppoatm.c |
| 38 | +++ b/net/atm/pppoatm.c |
| 39 | @@ -62,12 +62,25 @@ struct pppoatm_vcc { |
| 40 | void (*old_pop)(struct atm_vcc *, struct sk_buff *); |
| 41 | /* keep old push/pop for detaching */ |
| 42 | enum pppoatm_encaps encaps; |
| 43 | + atomic_t inflight; |
| 44 | + unsigned long blocked; |
| 45 | int flags; /* SC_COMP_PROT - compress protocol */ |
| 46 | struct ppp_channel chan; /* interface to generic ppp layer */ |
| 47 | struct tasklet_struct wakeup_tasklet; |
| 48 | }; |
| 49 | |
| 50 | /* |
| 51 | + * We want to allow two packets in the queue. The one that's currently in |
| 52 | + * flight, and *one* queued up ready for the ATM device to send immediately |
| 53 | + * from its TX done IRQ. We want to be able to use atomic_inc_not_zero(), so |
| 54 | + * inflight == -2 represents an empty queue, -1 one packet, and zero means |
| 55 | + * there are two packets in the queue. |
| 56 | + */ |
| 57 | +#define NONE_INFLIGHT -2 |
| 58 | + |
| 59 | +#define BLOCKED 0 |
| 60 | + |
| 61 | +/* |
| 62 | * Header used for LLC Encapsulated PPP (4 bytes) followed by the LCP protocol |
| 63 | * ID (0xC021) used in autodetection |
| 64 | */ |
| 65 | @@ -102,16 +115,30 @@ static void pppoatm_wakeup_sender(unsign |
| 66 | static void pppoatm_pop(struct atm_vcc *atmvcc, struct sk_buff *skb) |
| 67 | { |
| 68 | struct pppoatm_vcc *pvcc = atmvcc_to_pvcc(atmvcc); |
| 69 | + |
| 70 | pvcc->old_pop(atmvcc, skb); |
| 71 | + atomic_dec(&pvcc->inflight); |
| 72 | + |
| 73 | /* |
| 74 | - * We don't really always want to do this since it's |
| 75 | - * really inefficient - it would be much better if we could |
| 76 | - * test if we had actually throttled the generic layer. |
| 77 | - * Unfortunately then there would be a nasty SMP race where |
| 78 | - * we could clear that flag just as we refuse another packet. |
| 79 | - * For now we do the safe thing. |
| 80 | + * We always used to run the wakeup tasklet unconditionally here, for |
| 81 | + * fear of race conditions where we clear the BLOCKED flag just as we |
| 82 | + * refuse another packet in pppoatm_send(). This was quite inefficient. |
| 83 | + * |
| 84 | + * In fact it's OK. The PPP core will only ever call pppoatm_send() |
| 85 | + * while holding the channel->downl lock. And ppp_output_wakeup() as |
| 86 | + * called by the tasklet will *also* grab that lock. So even if another |
| 87 | + * CPU is in pppoatm_send() right now, the tasklet isn't going to race |
| 88 | + * with it. The wakeup *will* happen after the other CPU is safely out |
| 89 | + * of pppoatm_send() again. |
| 90 | + * |
| 91 | + * So if the CPU in pppoatm_send() has already set the BLOCKED bit and |
| 92 | + * it about to return, that's fine. We trigger a wakeup which will |
| 93 | + * happen later. And if the CPU in pppoatm_send() *hasn't* set the |
| 94 | + * BLOCKED bit yet, that's fine too because of the double check in |
| 95 | + * pppoatm_may_send() which is commented there. |
| 96 | */ |
| 97 | - tasklet_schedule(&pvcc->wakeup_tasklet); |
| 98 | + if (test_and_clear_bit(BLOCKED, &pvcc->blocked)) |
| 99 | + tasklet_schedule(&pvcc->wakeup_tasklet); |
| 100 | } |
| 101 | |
| 102 | /* |
| 103 | @@ -184,6 +211,51 @@ error: |
| 104 | ppp_input_error(&pvcc->chan, 0); |
| 105 | } |
| 106 | |
| 107 | +static inline int pppoatm_may_send(struct pppoatm_vcc *pvcc, int size) |
| 108 | +{ |
| 109 | + /* |
| 110 | + * It's not clear that we need to bother with using atm_may_send() |
| 111 | + * to check we don't exceed sk->sk_sndbuf. If userspace sets a |
| 112 | + * value of sk_sndbuf which is lower than the MTU, we're going to |
| 113 | + * block for ever. But the code always did that before we introduced |
| 114 | + * the packet count limit, so... |
| 115 | + */ |
| 116 | + if (atm_may_send(pvcc->atmvcc, size) && |
| 117 | + atomic_inc_not_zero_hint(&pvcc->inflight, NONE_INFLIGHT)) |
| 118 | + return 1; |
| 119 | + |
| 120 | + /* |
| 121 | + * We use test_and_set_bit() rather than set_bit() here because |
| 122 | + * we need to ensure there's a memory barrier after it. The bit |
| 123 | + * *must* be set before we do the atomic_inc() on pvcc->inflight. |
| 124 | + * There's no smp_mb__after_set_bit(), so it's this or abuse |
| 125 | + * smp_mb__after_clear_bit(). |
| 126 | + */ |
| 127 | + test_and_set_bit(BLOCKED, &pvcc->blocked); |
| 128 | + |
| 129 | + /* |
| 130 | + * We may have raced with pppoatm_pop(). If it ran for the |
| 131 | + * last packet in the queue, *just* before we set the BLOCKED |
| 132 | + * bit, then it might never run again and the channel could |
| 133 | + * remain permanently blocked. Cope with that race by checking |
| 134 | + * *again*. If it did run in that window, we'll have space on |
| 135 | + * the queue now and can return success. It's harmless to leave |
| 136 | + * the BLOCKED flag set, since it's only used as a trigger to |
| 137 | + * run the wakeup tasklet. Another wakeup will never hurt. |
| 138 | + * If pppoatm_pop() is running but hasn't got as far as making |
| 139 | + * space on the queue yet, then it hasn't checked the BLOCKED |
| 140 | + * flag yet either, so we're safe in that case too. It'll issue |
| 141 | + * an "immediate" wakeup... where "immediate" actually involves |
| 142 | + * taking the PPP channel's ->downl lock, which is held by the |
| 143 | + * code path that calls pppoatm_send(), and is thus going to |
| 144 | + * wait for us to finish. |
| 145 | + */ |
| 146 | + if (atm_may_send(pvcc->atmvcc, size) && |
| 147 | + atomic_inc_not_zero(&pvcc->inflight)) |
| 148 | + return 1; |
| 149 | + |
| 150 | + return 0; |
| 151 | +} |
| 152 | /* |
| 153 | * Called by the ppp_generic.c to send a packet - returns true if packet |
| 154 | * was accepted. If we return false, then it's our job to call |
| 155 | @@ -207,7 +279,7 @@ static int pppoatm_send(struct ppp_chann |
| 156 | struct sk_buff *n; |
| 157 | n = skb_realloc_headroom(skb, LLC_LEN); |
| 158 | if (n != NULL && |
| 159 | - !atm_may_send(pvcc->atmvcc, n->truesize)) { |
| 160 | + !pppoatm_may_send(pvcc, n->truesize)) { |
| 161 | kfree_skb(n); |
| 162 | goto nospace; |
| 163 | } |
| 164 | @@ -215,12 +287,12 @@ static int pppoatm_send(struct ppp_chann |
| 165 | skb = n; |
| 166 | if (skb == NULL) |
| 167 | return DROP_PACKET; |
| 168 | - } else if (!atm_may_send(pvcc->atmvcc, skb->truesize)) |
| 169 | + } else if (!pppoatm_may_send(pvcc, skb->truesize)) |
| 170 | goto nospace; |
| 171 | memcpy(skb_push(skb, LLC_LEN), pppllc, LLC_LEN); |
| 172 | break; |
| 173 | case e_vc: |
| 174 | - if (!atm_may_send(pvcc->atmvcc, skb->truesize)) |
| 175 | + if (!pppoatm_may_send(pvcc, skb->truesize)) |
| 176 | goto nospace; |
| 177 | break; |
| 178 | case e_autodetect: |
| 179 | @@ -285,6 +357,9 @@ static int pppoatm_assign_vcc(struct atm |
| 180 | if (pvcc == NULL) |
| 181 | return -ENOMEM; |
| 182 | pvcc->atmvcc = atmvcc; |
| 183 | + |
| 184 | + /* Maximum is zero, so that we can use atomic_inc_not_zero() */ |
| 185 | + atomic_set(&pvcc->inflight, NONE_INFLIGHT); |
| 186 | pvcc->old_push = atmvcc->push; |
| 187 | pvcc->old_pop = atmvcc->pop; |
| 188 | pvcc->encaps = (enum pppoatm_encaps) be.encaps; |
| 189 | |
