summaryrefslogtreecommitdiffstats
path: root/lib/pullwr.c
blob: 55ff4df011c06790ad84cbd73a117c575c8ce240 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
// SPDX-License-Identifier: GPL-2.0-or-later
/*
 * Pull-driven write event handler
 * Copyright (C) 2019  David Lamparter
 */

#include "zebra.h"

#include "pullwr.h"
#include "memory.h"
#include "monotime.h"

/* defaults */
#define PULLWR_THRESH	16384	/* size at which we start to call write() */
#define PULLWR_MAXSPIN	2500	/* max µs to spend grabbing more data */

struct pullwr {
	int fd;
	struct thread_master *tm;
	/* writer == NULL <=> we're idle */
	struct event *writer;

	void *arg;
	void (*fill)(void *, struct pullwr *);
	void (*err)(void *, struct pullwr *, bool);

	/* ring buffer (although it's "un-ringed" on resizing, it WILL wrap
	 * around if data is trickling in while keeping it at a constant size)
	 */
	size_t bufsz, valid, pos;
	uint64_t total_written;
	char *buffer;

	size_t thresh;		/* PULLWR_THRESH */
	int64_t maxspin;	/* PULLWR_MAXSPIN */
};

DEFINE_MTYPE_STATIC(LIB, PULLWR_HEAD, "pull-driven write controller");
DEFINE_MTYPE_STATIC(LIB, PULLWR_BUF,  "pull-driven write buffer");

static void pullwr_run(struct event *t);

struct pullwr *_pullwr_new(struct thread_master *tm, int fd,
		void *arg,
		void (*fill)(void *, struct pullwr *),
		void (*err)(void *, struct pullwr *, bool))
{
	struct pullwr *pullwr;

	pullwr = XCALLOC(MTYPE_PULLWR_HEAD, sizeof(*pullwr));
	pullwr->fd = fd;
	pullwr->tm = tm;
	pullwr->arg = arg;
	pullwr->fill = fill;
	pullwr->err = err;

	pullwr->thresh = PULLWR_THRESH;
	pullwr->maxspin = PULLWR_MAXSPIN;

	return pullwr;
}

void pullwr_del(struct pullwr *pullwr)
{
	THREAD_OFF(pullwr->writer);

	XFREE(MTYPE_PULLWR_BUF, pullwr->buffer);
	XFREE(MTYPE_PULLWR_HEAD, pullwr);
}

void pullwr_cfg(struct pullwr *pullwr, int64_t max_spin_usec,
		size_t write_threshold)
{
	pullwr->maxspin = max_spin_usec ?: PULLWR_MAXSPIN;
	pullwr->thresh = write_threshold ?: PULLWR_THRESH;
}

void pullwr_bump(struct pullwr *pullwr)
{
	if (pullwr->writer)
		return;

	event_add_timer(pullwr->tm, pullwr_run, pullwr, 0, &pullwr->writer);
}

static size_t pullwr_iov(struct pullwr *pullwr, struct iovec *iov)
{
	size_t len1;

	if (pullwr->valid == 0)
		return 0;

	if (pullwr->pos + pullwr->valid <= pullwr->bufsz) {
		iov[0].iov_base = pullwr->buffer + pullwr->pos;
		iov[0].iov_len = pullwr->valid;
		return 1;
	}

	len1 = pullwr->bufsz - pullwr->pos;

	iov[0].iov_base = pullwr->buffer + pullwr->pos;
	iov[0].iov_len = len1;
	iov[1].iov_base = pullwr->buffer;
	iov[1].iov_len = pullwr->valid - len1;
	return 2;
}

static void pullwr_resize(struct pullwr *pullwr, size_t need)
{
	struct iovec iov[2];
	size_t niov, newsize;
	char *newbuf;

	/* the buffer is maintained at pullwr->thresh * 2 since we'll be
	 * trying to fill it as long as it's anywhere below pullwr->thresh.
	 * That means we frequently end up a little short of it and then write
	 * something that goes over the threshold.  So, just use double.
	 */
	if (need) {
		/* resize up */
		if (pullwr->bufsz - pullwr->valid >= need)
			return;

		newsize = MAX((pullwr->valid + need) * 2, pullwr->thresh * 2);
		newbuf = XMALLOC(MTYPE_PULLWR_BUF, newsize);
	} else if (!pullwr->valid) {
		/* resize down, buffer empty */
		newsize = 0;
		newbuf = NULL;
	} else {
		/* resize down */
		if (pullwr->bufsz - pullwr->valid < pullwr->thresh)
			return;
		newsize = MAX(pullwr->valid, pullwr->thresh * 2);
		newbuf = XMALLOC(MTYPE_PULLWR_BUF, newsize);
	}

	niov = pullwr_iov(pullwr, iov);
	if (niov >= 1) {
		memcpy(newbuf, iov[0].iov_base, iov[0].iov_len);
		if (niov >= 2)
			memcpy(newbuf + iov[0].iov_len,
				iov[1].iov_base, iov[1].iov_len);
	}

	XFREE(MTYPE_PULLWR_BUF, pullwr->buffer);
	pullwr->buffer = newbuf;
	pullwr->bufsz = newsize;
	pullwr->pos = 0;
}

void pullwr_write(struct pullwr *pullwr, const void *data, size_t len)
{
	pullwr_resize(pullwr, len);

	if (pullwr->pos + pullwr->valid > pullwr->bufsz) {
		size_t pos;

		pos = (pullwr->pos + pullwr->valid) % pullwr->bufsz;
		memcpy(pullwr->buffer + pos, data, len);
	} else {
		size_t max1, len1;
		max1 = pullwr->bufsz - (pullwr->pos + pullwr->valid);
		max1 = MIN(max1, len);

		memcpy(pullwr->buffer + pullwr->pos + pullwr->valid,
				data, max1);
		len1 = len - max1;

		if (len1)
			memcpy(pullwr->buffer, (char *)data + max1, len1);

	}
	pullwr->valid += len;

	pullwr_bump(pullwr);
}

static void pullwr_run(struct event *t)
{
	struct pullwr *pullwr = THREAD_ARG(t);
	struct iovec iov[2];
	size_t niov, lastvalid;
	ssize_t nwr;
	struct timeval t0;
	bool maxspun = false;

	monotime(&t0);

	do {
		lastvalid = pullwr->valid - 1;
		while (pullwr->valid < pullwr->thresh
				&& pullwr->valid != lastvalid
				&& !maxspun) {
			lastvalid = pullwr->valid;
			pullwr->fill(pullwr->arg, pullwr);

			/* check after doing at least one fill() call so we
			 * don't spin without making progress on slow boxes
			 */
			if (!maxspun && monotime_since(&t0, NULL)
					>= pullwr->maxspin)
				maxspun = true;
		}

		if (pullwr->valid == 0) {
			/* we made a fill() call above that didn't feed any
			 * data in, and we have nothing more queued, so we go
			 * into idle, i.e. no calling event_add_write()
			 */
			pullwr_resize(pullwr, 0);
			return;
		}

		niov = pullwr_iov(pullwr, iov);
		assert(niov);

		nwr = writev(pullwr->fd, iov, niov);
		if (nwr < 0) {
			if (errno == EAGAIN || errno == EWOULDBLOCK)
				break;
			pullwr->err(pullwr->arg, pullwr, false);
			return;
		}

		if (nwr == 0) {
			pullwr->err(pullwr->arg, pullwr, true);
			return;
		}

		pullwr->total_written += nwr;
		pullwr->valid -= nwr;
		pullwr->pos += nwr;
		pullwr->pos %= pullwr->bufsz;
	} while (pullwr->valid == 0 && !maxspun);
	/* pullwr->valid != 0 implies we did an incomplete write, i.e. socket
	 * is full and we go wait until it's available for writing again.
	 */

	event_add_write(pullwr->tm, pullwr_run, pullwr, pullwr->fd,
			&pullwr->writer);

	/* if we hit the time limit, just keep the buffer, we'll probably need
	 * it anyway & another run is already coming up.
	 */
	if (!maxspun)
		pullwr_resize(pullwr, 0);
}

void pullwr_stats(struct pullwr *pullwr, uint64_t *total_written,
		  size_t *pending, size_t *kernel_pending)
{
	int tmp;

	*total_written = pullwr->total_written;
	*pending = pullwr->valid;

	if (ioctl(pullwr->fd, TIOCOUTQ, &tmp) != 0)
		tmp = 0;
	*kernel_pending = tmp;
}