wireproto: add streams to frame-based protocol
Previously, the frame-based protocol was just a series of frames,
with each frame associated with a request ID.
In order to scale the protocol, we'll want to enable the use of
compression. While it is possible to enable compression at the
socket/pipe level, this has its disadvantages. The big one is it
undermines the point of frames being standalone, atomic units that
can be read and written: if you add compression above the framing
protocol, you are back to having a stream-based protocol as opposed
to something frame-based.
So in order to preserve frames, compression needs to occur at
the frame payload level.
Compressing each frame's payload individually will limit compression
ratios because the window size of the compressor will be limited
by the max frame size, which is 32-64kb as currently defined. It
will also add CPU overhead, as it is more efficient for compressors
to operate on fewer, larger blocks of data than more, smaller blocks.
So compressing each frame independently is out.
This means we need to compress each frame's payload as if it is part
of a larger stream.
The simplest approach is to have 1 stream per connection. This
could certainly work. However, it has disadvantages (documented below).
We could also have 1 stream per RPC/command invocation. (This is the
model HTTP/2 goes with.) This also has disadvantages.
The main disadvantage to one global stream is that it has the very
real potential to create CPU bottlenecks doing compression. Networks
are only getting faster and the performance of single CPU cores has
been relatively flat. Newer compression formats like zstandard offer
better CPU cycle efficiency than predecessors like zlib. But it still
all too common to saturate your CPU with compression overhead long
before you saturate the network pipe.
The main disadvantage with streams per request is that you can't
reap the benefits of the compression context for multiple requests.
For example, if you send 1000 RPC requests (or HTTP/2 requests for
that matter), the response to each would have its own compression
context. The overall size of the raw responses would be larger because
compression contexts wouldn't be able to reference data from another
request or response.
The approach for streams as implemented in this commit is to support
N streams per connection and for streams to potentially span requests
and responses. As explained by the added internals docs, this
facilitates servers and clients delegating independent streams and
compression to independent threads / CPU cores. This helps alleviate
the CPU bottleneck of compression. This design also allows compression
contexts to be reused across requests/responses. This can result in
improved compression ratios and less overhead for compressors and
decompressors having to build new contexts.
Another feature that was defined was the ability for individual frames
within a stream to declare whether that individual frame's payload
uses the content encoding (read: compression) defined by the stream.
The idea here is that some servers may serve data from a combination
of caches and dynamic resolution. Data coming from caches may be
pre-compressed. We want to facilitate servers being able to essentially
stream bytes from caches to the wire with minimal overhead. Being
able to mix and match with frames are compressed within a stream
enables these types of advanced server functionality.
This commit defines the new streams mechanism. Basic code for
supporting streams in frames has been added. But that code is
seriously lacking and doesn't fully conform to the defined protocol.
For example, we don't close any streams. And support for content
encoding within streams is not yet implemented. The change was
rather invasive and I didn't think it would be reasonable to implement
the entire feature in a single commit.
For the record, I would have loved to reuse an existing multiplexing
protocol to build the new wire protocol on top of. However, I couldn't
find a protocol that offers the performance and scaling characteristics
that I desired. Namely, it should support multiple compression
contexts to facilitate scaling out to multiple CPU cores and
compression contexts should be able to live longer than single RPC
requests. HTTP/2 *almost* fits the bill. But the semantics of HTTP
message exchange state that streams can only live for a single
request-response. We /could/ tunnel on top of HTTP/2 streams and
frames with HEADER and DATA frames. But there's no guarantee that
HTTP/2 libraries and proxies would allow us to use HTTP/2 streams
and frames without the HTTP message exchange semantics defined in
RFC 7540 Section 8. Other RPC protocols like gRPC tunnel are built
on top of HTTP/2 and thus preserve its semantics of stream per
RPC invocation. Even QUIC does this. We could attempt to invent a
higher-level stream that spans HTTP/2 streams. But this would be
violating HTTP/2 because there is no guarantee that HTTP/2 streams
are routed to the same server. The best we can do - which is what
this protocol does - is shoehorn all request and response data into
a single HTTP message and create streams within. At that point, we've
defined a Content-Type in HTTP parlance. It just so happens our
media type can also work as a standalone, stream-based protocol,
without leaning on HTTP or similar protocol.
Differential Revision: https://phab.mercurial-scm.org/D2907
/*
bdiff.c - efficient binary diff extension for Mercurial
Copyright 2005, 2006 Matt Mackall <mpm@selenic.com>
This software may be used and distributed according to the terms of
the GNU General Public License, incorporated herein by reference.
Based roughly on Python difflib
*/
#include <limits.h>
#include <stdlib.h>
#include <string.h>
#include "bdiff.h"
#include "bitmanipulation.h"
#include "compat.h"
/* Hash implementation from diffutils */
#define ROL(v, n) ((v) << (n) | (v) >> (sizeof(v) * CHAR_BIT - (n)))
#define HASH(h, c) ((c) + ROL(h, 7))
struct pos {
int pos, len;
};
int bdiff_splitlines(const char *a, ssize_t len, struct bdiff_line **lr)
{
unsigned hash;
int i;
const char *p, *b = a;
const char *const plast = a + len - 1;
struct bdiff_line *l;
/* count the lines */
i = 1; /* extra line for sentinel */
for (p = a; p < plast; p++)
if (*p == '\n')
i++;
if (p == plast)
i++;
*lr = l = (struct bdiff_line *)calloc(i, sizeof(struct bdiff_line));
if (!l)
return -1;
/* build the line array and calculate hashes */
hash = 0;
for (p = a; p < plast; p++) {
hash = HASH(hash, *p);
if (*p == '\n') {
l->hash = hash;
hash = 0;
l->len = p - b + 1;
l->l = b;
l->n = INT_MAX;
l++;
b = p + 1;
}
}
if (p == plast) {
hash = HASH(hash, *p);
l->hash = hash;
l->len = p - b + 1;
l->l = b;
l->n = INT_MAX;
l++;
}
/* set up a sentinel */
l->hash = 0;
l->len = 0;
l->l = a + len;
return i - 1;
}
static inline int cmp(struct bdiff_line *a, struct bdiff_line *b)
{
return a->hash != b->hash || a->len != b->len ||
memcmp(a->l, b->l, a->len);
}
static int equatelines(struct bdiff_line *a, int an, struct bdiff_line *b,
int bn)
{
int i, j, buckets = 1, t, scale;
struct pos *h = NULL;
/* build a hash table of the next highest power of 2 */
while (buckets < bn + 1)
buckets *= 2;
/* try to allocate a large hash table to avoid collisions */
for (scale = 4; scale; scale /= 2) {
h = (struct pos *)calloc(buckets, scale * sizeof(struct pos));
if (h)
break;
}
if (!h)
return 0;
buckets = buckets * scale - 1;
/* clear the hash table */
for (i = 0; i <= buckets; i++) {
h[i].pos = -1;
h[i].len = 0;
}
/* add lines to the hash table chains */
for (i = 0; i < bn; i++) {
/* find the equivalence class */
for (j = b[i].hash & buckets; h[j].pos != -1;
j = (j + 1) & buckets)
if (!cmp(b + i, b + h[j].pos))
break;
/* add to the head of the equivalence class */
b[i].n = h[j].pos;
b[i].e = j;
h[j].pos = i;
h[j].len++; /* keep track of popularity */
}
/* compute popularity threshold */
t = (bn >= 31000) ? bn / 1000 : 1000000 / (bn + 1);
/* match items in a to their equivalence class in b */
for (i = 0; i < an; i++) {
/* find the equivalence class */
for (j = a[i].hash & buckets; h[j].pos != -1;
j = (j + 1) & buckets)
if (!cmp(a + i, b + h[j].pos))
break;
a[i].e = j; /* use equivalence class for quick compare */
if (h[j].len <= t)
a[i].n = h[j].pos; /* point to head of match list */
else
a[i].n = -1; /* too popular */
}
/* discard hash tables */
free(h);
return 1;
}
static int longest_match(struct bdiff_line *a, struct bdiff_line *b,
struct pos *pos, int a1, int a2, int b1, int b2,
int *omi, int *omj)
{
int mi = a1, mj = b1, mk = 0, i, j, k, half, bhalf;
/* window our search on large regions to better bound
worst-case performance. by choosing a window at the end, we
reduce skipping overhead on the b chains. */
if (a2 - a1 > 30000)
a1 = a2 - 30000;
half = (a1 + a2 - 1) / 2;
bhalf = (b1 + b2 - 1) / 2;
for (i = a1; i < a2; i++) {
/* skip all lines in b after the current block */
for (j = a[i].n; j >= b2; j = b[j].n)
;
/* loop through all lines match a[i] in b */
for (; j >= b1; j = b[j].n) {
/* does this extend an earlier match? */
for (k = 1; j - k >= b1 && i - k >= a1; k++) {
/* reached an earlier match? */
if (pos[j - k].pos == i - k) {
k += pos[j - k].len;
break;
}
/* previous line mismatch? */
if (a[i - k].e != b[j - k].e)
break;
}
pos[j].pos = i;
pos[j].len = k;
/* best match so far? we prefer matches closer
to the middle to balance recursion */
if (k > mk) {
/* a longer match */
mi = i;
mj = j;
mk = k;
} else if (k == mk) {
if (i > mi && i <= half && j > b1) {
/* same match but closer to half */
mi = i;
mj = j;
} else if (i == mi && (mj > bhalf || i == a1)) {
/* same i but best earlier j */
mj = j;
}
}
}
}
if (mk) {
mi = mi - mk + 1;
mj = mj - mk + 1;
}
/* expand match to include subsequent popular lines */
while (mi + mk < a2 && mj + mk < b2 && a[mi + mk].e == b[mj + mk].e)
mk++;
*omi = mi;
*omj = mj;
return mk;
}
static struct bdiff_hunk *recurse(struct bdiff_line *a, struct bdiff_line *b,
struct pos *pos, int a1, int a2, int b1,
int b2, struct bdiff_hunk *l)
{
int i, j, k;
while (1) {
/* find the longest match in this chunk */
k = longest_match(a, b, pos, a1, a2, b1, b2, &i, &j);
if (!k)
return l;
/* and recurse on the remaining chunks on either side */
l = recurse(a, b, pos, a1, i, b1, j, l);
if (!l)
return NULL;
l->next =
(struct bdiff_hunk *)malloc(sizeof(struct bdiff_hunk));
if (!l->next)
return NULL;
l = l->next;
l->a1 = i;
l->a2 = i + k;
l->b1 = j;
l->b2 = j + k;
l->next = NULL;
/* tail-recursion didn't happen, so do equivalent iteration */
a1 = i + k;
b1 = j + k;
}
}
int bdiff_diff(struct bdiff_line *a, int an, struct bdiff_line *b, int bn,
struct bdiff_hunk *base)
{
struct bdiff_hunk *curr;
struct pos *pos;
int t, count = 0;
/* allocate and fill arrays */
t = equatelines(a, an, b, bn);
pos = (struct pos *)calloc(bn ? bn : 1, sizeof(struct pos));
if (pos && t) {
/* generate the matching block list */
curr = recurse(a, b, pos, 0, an, 0, bn, base);
if (!curr)
return -1;
/* sentinel end hunk */
curr->next =
(struct bdiff_hunk *)malloc(sizeof(struct bdiff_hunk));
if (!curr->next)
return -1;
curr = curr->next;
curr->a1 = curr->a2 = an;
curr->b1 = curr->b2 = bn;
curr->next = NULL;
}
free(pos);
/* normalize the hunk list, try to push each hunk towards the end */
for (curr = base->next; curr; curr = curr->next) {
struct bdiff_hunk *next = curr->next;
if (!next)
break;
if (curr->a2 == next->a1 || curr->b2 == next->b1)
while (curr->a2 < an && curr->b2 < bn &&
next->a1 < next->a2 && next->b1 < next->b2 &&
!cmp(a + curr->a2, b + curr->b2)) {
curr->a2++;
next->a1++;
curr->b2++;
next->b1++;
}
}
for (curr = base->next; curr; curr = curr->next)
count++;
return count;
}
void bdiff_freehunks(struct bdiff_hunk *l)
{
struct bdiff_hunk *n;
for (; l; l = n) {
n = l->next;
free(l);
}
}