2 * Copyright (c) 2016-2018 Netflix, Inc.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #ifndef __tcp_hpts_h__
29 #define __tcp_hpts_h__
32 * The hpts uses a 102400 wheel. The wheel
33 * defines the time in 10 usec increments (102400 x 10).
34 * This gives a range of 10usec - 1024ms to place
35 * an entry within. If the user requests more than
36 * 1.024 second, a remaineder is attached and the hpts
37 * when seeing the remainder will re-insert the
38 * inpcb forward in time from where it is until
39 * the remainder is zero.
42 #define NUM_OF_HPTSI_SLOTS 102400
44 TAILQ_HEAD(hptsh, inpcb);
46 /* Number of useconds in a hpts tick */
47 #define HPTS_TICKS_PER_USEC 10
48 #define HPTS_MS_TO_SLOTS(x) ((x * 100) + 1)
49 #define HPTS_USEC_TO_SLOTS(x) ((x+9) /10)
50 #define HPTS_USEC_IN_SEC 1000000
51 #define HPTS_MSEC_IN_SEC 1000
52 #define HPTS_USEC_IN_MSEC 1000
55 uint32_t p_hpts_active; /* bbr->flex7 x */
56 uint32_t p_nxt_slot; /* bbr->flex1 x */
57 uint32_t p_cur_slot; /* bbr->flex2 x */
58 uint32_t p_prev_slot; /* bbr->delivered */
59 uint32_t p_runningtick; /* bbr->inflight */
60 uint32_t slot_req; /* bbr->flex3 x */
61 uint32_t inp_hptsslot; /* bbr->flex4 x */
62 uint32_t slot_remaining; /* bbr->flex5 x */
63 uint32_t have_slept; /* bbr->epoch x */
64 uint32_t hpts_sleep_time; /* bbr->applimited x */
65 uint32_t yet_to_sleep; /* bbr->lt_epoch x */
66 uint32_t need_new_to; /* bbr->flex6 x */
67 uint32_t wheel_tick; /* bbr->bw_inuse x */
68 uint32_t maxticks; /* bbr->delRate x */
69 uint32_t wheel_cts; /* bbr->rttProp x */
70 int32_t co_ret; /* bbr->pkts_out x */
71 uint32_t p_curtick; /* upper bbr->cur_del_rate */
72 uint32_t p_lasttick; /* lower bbr->cur_del_rate */
73 uint8_t p_on_min_sleep; /* bbr->flex8 x */
76 /* Magic flags to tell whats cooking on the pacing wheel */
77 #define PACE_TMR_DELACK 0x01 /* Delayed ack timer running */
78 #define PACE_TMR_RACK 0x02 /* RACK timer running */
79 #define PACE_TMR_TLP 0x04 /* TLP timer running */
80 #define PACE_TMR_RXT 0x08 /* Retransmit timer running */
81 #define PACE_TMR_PERSIT 0x10 /* Persists timer running */
82 #define PACE_TMR_KEEP 0x20 /* Keep alive timer running */
83 #define PACE_PKT_OUTPUT 0x40 /* Output Packets being paced */
84 #define PACE_TMR_MASK (PACE_TMR_KEEP|PACE_TMR_PERSIT|PACE_TMR_RXT|PACE_TMR_TLP|PACE_TMR_RACK|PACE_TMR_DELACK)
87 /* Each hpts has its own p_mtx which is used for locking */
88 struct tcp_hpts_entry {
90 struct mtx p_mtx; /* Mutex for hpts */
91 uint16_t p_hpts_active; /* Flag that says hpts is awake */
92 uint8_t p_hpts_wake_scheduled; /* Have we scheduled a wakeup? */
93 uint8_t p_wheel_complete; /* have we completed the wheel arc walk? */
94 uint32_t p_curtick; /* Tick in 10 us the hpts is going to */
95 uint32_t p_runningtick; /* Current tick we are at if we are running */
96 uint32_t p_prev_slot; /* Previous slot we were on */
97 uint32_t p_cur_slot; /* Current slot in wheel hpts is draining */
98 uint32_t p_nxt_slot; /* The next slot outside the current range of
99 * slots that the hpts is running on. */
100 int32_t p_on_queue_cnt; /* Count on queue in this hpts */
101 uint32_t p_lasttick; /* Last tick before the current one */
102 uint8_t p_direct_wake :1, /* boolean */
103 p_on_min_sleep:1, /* boolean */
105 uint8_t p_fill[3]; /* Fill to 32 bits */
106 /* Cache line 0x40 */
108 struct hptsh p_input; /* For the tcp-input runner */
110 struct hptsh *p_hptss;
111 int32_t p_on_inqueue_cnt; /* Count on input queue in this hpts */
112 uint32_t hit_no_enobuf;
113 uint32_t p_dyn_adjust;
114 uint32_t p_hpts_sleep_time; /* Current sleep interval having a max
116 uint32_t overidden_sleep; /* what was overrided by min-sleep for logging */
117 uint32_t saved_lasttick; /* for logging */
118 uint32_t saved_curtick; /* for logging */
119 uint32_t saved_curslot; /* for logging */
120 uint32_t saved_prev_slot; /* for logging */
121 uint32_t p_delayed_by; /* How much were we delayed by */
122 /* Cache line 0x80 */
123 struct sysctl_ctx_list hpts_ctx;
124 struct sysctl_oid *hpts_root;
125 struct intr_event *ie;
127 uint16_t p_num; /* The hpts number one per cpu */
128 uint16_t p_cpu; /* The hpts CPU */
129 /* There is extra space in here */
130 /* Cache line 0x100 */
131 struct callout co __aligned(CACHE_LINE_SIZE);
132 } __aligned(CACHE_LINE_SIZE);
135 struct proc *rp_proc; /* Process structure for hpts */
136 struct tcp_hpts_entry **rp_ent; /* Array of hptss */
137 uint32_t rp_num_hptss; /* Number of hpts threads */
142 #define HPTS_REMOVE_INPUT 0x01
143 #define HPTS_REMOVE_OUTPUT 0x02
144 #define HPTS_REMOVE_ALL (HPTS_REMOVE_INPUT | HPTS_REMOVE_OUTPUT)
147 * When using the hpts, a TCP stack must make sure
148 * that once a INP_DROPPED flag is applied to a INP
149 * that it does not expect tcp_output() to ever be
150 * called by the hpts. The hpts will *not* call
151 * any output (or input) functions on a TCB that
152 * is in the DROPPED state.
154 * This implies final ACK's and RST's that might
155 * be sent when a TCB is still around must be
156 * sent from a routine like tcp_respond().
158 #define DEFAULT_MIN_SLEEP 250 /* How many usec's is default for hpts sleep
159 * this determines min granularity of the
160 * hpts. If 0, granularity is 10useconds at
161 * the cost of more CPU (context switching). */
163 #define HPTS_MTX_ASSERT(hpts) mtx_assert(&(hpts)->p_mtx, MA_OWNED)
164 struct tcp_hpts_entry *tcp_hpts_lock(struct inpcb *inp);
165 struct tcp_hpts_entry *tcp_input_lock(struct inpcb *inp);
166 int __tcp_queue_to_hpts_immediate(struct inpcb *inp, int32_t line);
167 #define tcp_queue_to_hpts_immediate(a)__tcp_queue_to_hpts_immediate(a, __LINE__)
169 struct tcp_hpts_entry *tcp_cur_hpts(struct inpcb *inp);
170 #define tcp_hpts_remove(a, b) __tcp_hpts_remove(a, b, __LINE__)
171 void __tcp_hpts_remove(struct inpcb *inp, int32_t flags, int32_t line);
174 * To insert a TCB on the hpts you *must* be holding the
175 * INP_WLOCK(). The hpts insert code will then acqurire
176 * the hpts's lock and insert the TCB on the requested
177 * slot possibly waking up the hpts if you are requesting
178 * a time earlier than what the hpts is sleeping to (if
179 * the hpts is sleeping). You may check the inp->inp_in_hpts
180 * flag without the hpts lock. The hpts is the only one
181 * that will clear this flag holding only the hpts lock. This
182 * means that in your tcp_output() routine when you test for
183 * it to be 1 (so you wont call output) it may be transitioning
184 * to 0 (by the hpts). That will be fine since that will just
185 * mean an extra call to tcp_output that most likely will find
186 * the call you executed (when the mis-match occured) will have
187 * put the TCB back on the hpts and it will return. If your
188 * call did not add it back to the hpts then you will either
189 * over-send or the cwnd will block you from sending more.
191 * Note you should also be holding the INP_WLOCK() when you
192 * call the remove from the hpts as well. Thoug usually
193 * you are either doing this from a timer, where you need
194 * that INP_WLOCK() or from destroying your TCB where again
195 * you should already have the INP_WLOCK().
197 uint32_t __tcp_hpts_insert(struct inpcb *inp, uint32_t slot, int32_t line);
198 #define tcp_hpts_insert(a, b) __tcp_hpts_insert(a, b, __LINE__)
201 tcp_hpts_insert_diag(struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag);
204 __tcp_queue_to_input_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line);
205 #define tcp_queue_to_input_locked(a, b) __tcp_queue_to_input_locked(a, b, __LINE__);
207 __tcp_queue_to_input(struct inpcb *inp, int32_t line);
208 #define tcp_queue_to_input(a) __tcp_queue_to_input(a, __LINE__)
210 uint16_t tcp_hpts_delayedby(struct inpcb *inp);
212 void __tcp_set_hpts(struct inpcb *inp, int32_t line);
213 #define tcp_set_hpts(a) __tcp_set_hpts(a, __LINE__)
215 void __tcp_set_inp_to_drop(struct inpcb *inp, uint16_t reason, int32_t line);
216 #define tcp_set_inp_to_drop(a, b) __tcp_set_inp_to_drop(a, b, __LINE__)
218 extern int32_t tcp_min_hptsi_time;
220 static __inline uint32_t
221 tcp_tv_to_hptstick(struct timeval *sv)
223 return ((sv->tv_sec * 100000) + (sv->tv_usec / 10));
226 static __inline uint32_t
227 tcp_gethptstick(struct timeval *sv)
234 return (tcp_tv_to_hptstick(sv));
237 static __inline uint32_t
238 tcp_tv_to_usectick(struct timeval *sv)
240 return ((uint32_t) ((sv->tv_sec * HPTS_USEC_IN_SEC) + sv->tv_usec));
243 static __inline uint32_t
244 tcp_tv_to_mssectick(struct timeval *sv)
246 return ((uint32_t) ((sv->tv_sec * HPTS_MSEC_IN_SEC) + (sv->tv_usec/HPTS_USEC_IN_MSEC)));
250 tcp_hpts_unlock(struct tcp_hpts_entry *hpts)
252 mtx_unlock(&hpts->p_mtx);
255 static __inline uint32_t
256 tcp_get_usecs(struct timeval *tv)
263 return (tcp_tv_to_usectick(tv));
267 #endif /* __tcp_hpts_h__ */