http: add support for HTTP 429 rate limit retries
Subsystems (2)
cover http: add support for HTTP 429 rate limit retries
This patch series adds support for handling HTTP 429 (Too Many Requests) responses in Git's HTTP client with automatic retry logic.
Git hosting services can implement rate limiting to protect their infrastructure. When these limits are reached, servers respond with HTTP 429 status codes, potentially including a Retry-After header to indicate when the client should retry. Currently, Git treats these responses as fatal errors, forcing users to manually retry their operations.
This series implements automatic retry support with three new configuration options:
* http.maxRetries: Controls the maximum number of retry attempts (default:
0, opt-in behavior)
* http.retryAfter: Provides a fallback delay when the server doesn't include a Retry-After header (default: -1, fail if no header)
* http.maxRetryTime: Sets an upper limit on any single retry delay (default: 300 seconds) to prevent indefinite blocking
The implementation includes:
Patch 1: Core HTTP 429 retry logic with support for RFC-compliant Retry-After headers (both delay-seconds and HTTP-date formats), comprehensive configuration options, and fail-fast behavior for excessive delays. Includes extensive test coverage.
Patch 2: Fixes a pre-existing memory leak in show_http_message() that became more visible with the new retry logic.
Patch 3: Adds trace2 instrumentation to enable monitoring and debugging of retry operations in production environments.
The retry behavior is disabled by default (maxRetries = 0), requiring explicit opt-in, ensuring backward compatibility while providing a robust solution for environments that need rate limit handling.
There was a previous attempt to add retry support [1], which was not merged. It had support for 50x status codes. Should they be supported here too?
[1] https://lore.kernel.org/git/[email protected]/
Vaidas Pilkauskas (3): http: add support for HTTP 429 rate limit retries
remote-curl: fix memory leak in show_http_message() http: add trace2 logging for retry operations
Documentation/config/http.adoc | 24 ++ http-push.c | 8 + http-walker.c | 5 + http.c | 171 ++++++++++++- http.h | 2 + remote-curl.c | 18 +- t/meson.build | 1 + t/t5584-http-429-retry.sh | 429 +++++++++++++++++++++++++++++++++ 8 files changed, 650 insertions(+), 8 deletions(-) create mode 100755 t/t5584-http-429-retry.sh
-- 2.50.1
Vaidas Pilkauskas (3):
http: add support for HTTP 429 rate limit retries
remote-curl: fix memory leak in show_http_message()
http: add trace2 logging for retry operations
Documentation/config/http.adoc | 24 ++ http-push.c | 8 + http-walker.c | 5 + http.c | 171 ++++++++++++- http.h | 2 + remote-curl.c | 18 +- t/meson.build | 1 + t/t5584-http-429-retry.sh | 429 +++++++++++++++++++++++++++++++++ 8 files changed, 650 insertions(+), 8 deletions(-) create mode 100755 t/t5584-http-429-retry.sh
base-commit: 6ab38b7e9cc7adafc304f3204616a4debd49c6e9
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-2008%2Fvaidas-shopify%2Fretry-after-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-2008/vaidas-shopify/retry-after-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/2008
Signature · 1 line
gitgitgadget
PATCH 1/3 http: add support for HTTP 429 rate limit retries
From: Vaidas Pilkauskas <[email protected]>
Add retry logic for HTTP 429 (Too Many Requests) responses to handle server-side rate limiting gracefully. When Git's HTTP client receives a 429 response, it can now automatically retry the request after an appropriate delay, respecting the server's rate limits.
The implementation supports the RFC-compliant Retry-After header in both delay-seconds (integer) and HTTP-date (RFC 2822) formats. If a past date is provided, Git retries immediately without waiting.
Retry behavior is controlled by three new configuration options:
* http.maxRetries: Maximum number of retry attempts (default: 0, meaning retries are disabled by default). Users must explicitly opt-in to retry behavior.
* http.retryAfter: Default delay in seconds when the server doesn't provide a Retry-After header (default: -1, meaning fail if no header is provided). This serves as a fallback mechanism.
* http.maxRetryTime: Maximum delay in seconds for a single retry (default: 300). If the server requests a delay exceeding this limit, Git fails immediately rather than waiting. This prevents indefinite blocking on unreasonable server requests.
All three options can be overridden via environment variables:GIT_HTTP_MAX_RETRIES, GIT_HTTP_RETRY_AFTER, and GIT_HTTP_MAX_RETRY_TIME.
The retry logic implements a fail-fast approach: if any delay (whether from server header or configuration) exceeds maxRetryTime, Git fails immediately with a clear error message rather than capping the delay. This provides better visibility into rate limiting issues.
The implementation includes extensive test coverage for basic retry behavior, Retry-After header formats (integer and HTTP-date), configuration combinations, maxRetryTime limits, invalid header handling, environment variable overrides, and edge cases.
--- Documentation/config/http.adoc | 24 ++ http-push.c | 8 + http-walker.c | 5 + http.c | 149 +++++++++++- http.h | 2 + remote-curl.c | 4 + t/meson.build | 1 + t/t5584-http-429-retry.sh | 429 +++++++++++++++++++++++++++++++++ 8 files changed, 618 insertions(+), 4 deletions(-) create mode 100755 t/t5584-http-429-retry.sh
Documentation/config/http.adoc +24 -0
@@ -315,6 +315,30 @@ http.keepAliveCount:: unset, curl's default value is used. Can be overridden by the `GIT_HTTP_KEEPALIVE_COUNT` environment variable. +http.retryAfter:: + Default wait time in seconds before retrying when a server returns + HTTP 429 (Too Many Requests) without a Retry-After header. If set + to -1 (the default), Git will fail immediately when encountering + a 429 response without a Retry-After header. When a Retry-After + header is present, its value takes precedence over this setting. + Can be overridden by the `GIT_HTTP_RETRY_AFTER` environment variable. + See also `http.maxRetries` and `http.maxRetryTime`. + +http.maxRetries:: + Maximum number of times to retry after receiving HTTP 429 (Too Many + Requests) responses. Set to 0 (the default) to disable retries. + Can be overridden by the `GIT_HTTP_MAX_RETRIES` environment variable. + See also `http.retryAfter` and `http.maxRetryTime`. + +http.maxRetryTime:: + Maximum time in seconds to wait for a single retry attempt when + handling HTTP 429 (Too Many Requests) responses. If the server + requests a delay (via Retry-After header) or if `http.retryAfter` + is configured with a value that exceeds this maximum, Git will fail + immediately rather than waiting. Default is 300 seconds (5 minutes). + Can be overridden by the `GIT_HTTP_MAX_RETRY_TIME` environment + variable. See also `http.retryAfter` and `http.maxRetries`. + http.noEPSV:: A boolean which disables using of EPSV ftp command by curl. This can be helpful with some "poor" ftp servers which don't
http-push.c +8 -0
@@ -716,6 +716,10 @@ static int fetch_indices(void)case HTTP_MISSING_TARGET : ret =0 ;break ; +case HTTP_RATE_LIMITED : +error ("rate limited by '%s', please try again later" , repo->url); + ret =-1 ; +break ;default : ret =-1 ; }
@@ -1548,6 +1552,10 @@ static int remote_exists(const char *path)case HTTP_MISSING_TARGET : ret =0 ;break ; +case HTTP_RATE_LIMITED : +error ("rate limited by '%s', please try again later" , url); + ret =-1 ; +break ;case HTTP_ERROR :error ("unable to access '%s': %s" , url, curl_errorstr);/* fallthrough */
http-walker.c +5 -0
@@ -414,6 +414,11 @@ static int fetch_indices(struct walker *walker, struct alt_base *repo) repo->got_indices =1 ; ret =0 ;break ; +case HTTP_RATE_LIMITED : +error ("rate limited by '%s', please try again later" , repo->base); + repo->got_indices =0 ; + ret =-1 ; +break ;default : repo->got_indices =0 ; ret =-1 ;
http.c +145 -4
239 lines not on this page · raw message
http.h +2 -0
@@ -20,6 +20,7 @@ struct slot_results {long http_code;long auth_avail;long http_connectcode; +long retry_after; };struct active_request_slot {
@@ -167,6 +168,7 @@ struct http_get_options {#define HTTP_REAUTH 4#define HTTP_NOAUTH 5#define HTTP_NOMATCHPUBLICKEY 6 +#define HTTP_RATE_LIMITED 7/* * Requests a URL and stores the result in a strbuf.
remote-curl.c +4 -0
@@ -529,6 +529,10 @@ static struct discovery *discover_refs(const char *service, int for_push)show_http_message (&type, &charset, &buffer);die (_ ("unable to access '%s' with http.pinnedPubkey configuration: %s" ),transport_anonymize_url (url.buf), curl_errorstr); +case HTTP_RATE_LIMITED : +show_http_message (&type, &charset, &buffer); +die (_ ("rate limited by '%s', please try again later" ), +transport_anonymize_url (url.buf));default :show_http_message (&type, &charset, &buffer); die(_("unable to access '%s': %s"),
t/meson.build +1 -0
@@ -698,6 +698,7 @@ integration_tests = [ 't5581-http-curl-verbose.sh', 't5582-fetch-negative-refspec.sh', 't5583-push-branches.sh', + 't5584-http-429-retry.sh', 't5600-clone-fail-cleanup.sh', 't5601-clone.sh', 't5602-clone-remote-exec.sh',
t/t5584-http-429-retry.sh new file +429 -0
429 lines not on this page · raw message
Signature · 1 line
gitgitgadget
PATCH 2/3 remote-curl: fix memory leak in show_http_message()
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
Fix a memory leak in show_http_message() that was triggered when displaying HTTP error messages before die(). The function would call strbuf_reencode() which modifies the caller's strbuf in place, allocating new memory for the re-encoded string. Since this function is only called immediately before die(), the allocated memory was never explicitly freed, causing leak detectors to report it.
The leak became visible when HTTP 429 rate limit retry support was added, which introduced the HTTP_RATE_LIMITED error case. However, the issue existed in pre-existing error paths as well (HTTP_MISSING_TARGET, HTTP_NOAUTH, HTTP_NOMATCHPUBLICKEY) - the new retry logic just made it more visible in tests because retries exercise the error paths more frequently.
The leak was detected by LeakSanitizer in t5584 tests that enable retries (maxRetries > 0). Tests with retries disabled passed because they took a different code path or timing.
Fix this by making show_http_message() work on a local copy of the message buffer instead of modifying the caller's buffer in place:
1. Create a local strbuf and copy the message into it
2. Perform re-encoding on the local copy if needed
3. Display the message from the local copy
4. Properly release the local copy before returning
This ensures all memory allocated by strbuf_reencode() is freed before the function returns, even though die() is called immediately after, eliminating the leak.
--- remote-curl.c | 14 ++++++++++---- 1 file changed, 10 insertions(+), 4 deletions(-)
remote-curl.c +10 -4
@@ -371,6 +371,7 @@ static int show_http_message(struct strbuf *type, struct strbuf *charset, struct strbuf *msg) {const char *p, *eol; +struct strbuf msgbuf =STRBUF_INIT ;/* * We only show text/plain parts, as other types are likely
@@ -378,19 +379,24 @@ static int show_http_message(struct strbuf *type, struct strbuf *charset,*/ if (strcmp (type->buf,"text/plain" ))return -1 ; + +strbuf_addbuf (&msgbuf, msg);if (charset->len) -strbuf_reencode (msg, charset->buf,get_log_output_encoding ()); +strbuf_reencode (&msgbuf, charset->buf,get_log_output_encoding ()); -strbuf_trim (msg); -if (!msg->len) +strbuf_trim (&msgbuf); +if (!msgbuf.len) { +strbuf_release (&msgbuf);return -1 ; + } - p = msg->buf; + p = msgbuf.buf;do { eol =strchrnul (p,'\n' );fprintf (stderr,"remote: %.*s\n" , (int )(eol - p), p); p = eol +1 ; }while (*eol); +strbuf_release (&msgbuf);return 0 ; }
Signature · 1 line
gitgitgadget
PATCH 3/3 http: add trace2 logging for retry operations
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
Add trace2 instrumentation to HTTP 429 retry operations to enable monitoring and debugging of rate limit scenarios in production environments.
The trace2 logging captures:
* Retry attempt numbers (http/429-retry-attempt) to track retry progression and identify how many attempts were needed
* Retry-After header values (http/429-retry-after) from server responses to understand server-requested delays
* Actual sleep durations (http/retry-sleep-seconds) within trace2 regions (http/retry-sleep) to measure time spent waiting
* Error conditions (http/429-error) such as "retries-exhausted", "exceeds-max-retry-time", "no-retry-after-config", and "config-exceeds-max-retry-time" for diagnosing failures
* Retry source (http/429-retry-source) indicating whether delay came from server header or config default
This instrumentation provides complete visibility into retry behavior, enabling operators to monitor rate limiting patterns, diagnose retry failures, and optimize retry configuration based on real-world data.
--- http.c | 40 +++++++++++++++++++++++++++++++--------- 1 file changed, 31 insertions(+), 9 deletions(-)
http.c +31 -9
@@ -23,6 +23,7 @@#include "odb.h" #include "tempfile.h" #include "date.h" +#include "trace2.h" static struct trace_key trace_curl =TRACE_KEY_INIT (CURL );static int trace_curl_data =1 ;
@@ -1944,6 +1945,8 @@ static int handle_curl_result(struct slot_results *results) }else if (results->http_code ==429 ) {/* Store the retry_after value for use in retry logic */ last_retry_after = results->retry_after; +trace2_data_intmax ("http" , the_repository,"http/429-retry-after" , + last_retry_after);return HTTP_RATE_LIMITED ; }else {if (results->http_connectcode ==407 )
@@ -2338,11 +2341,15 @@ static void sleep_for_retry(long retry_after)if (retry_after >0 ) {unsigned int remaining;warning (_ ("rate limited, waiting %ld seconds before retry" ), retry_after); +trace2_region_enter ("http" ,"retry-sleep" , the_repository); +trace2_data_intmax ("http" , the_repository,"http/retry-sleep-seconds" , + retry_after); remaining =sleep (retry_after);while (remaining >0 ) {/* Sleep was interrupted, continue sleeping */ remaining =sleep (remaining); } +trace2_region_leave ("http" ,"retry-sleep" , the_repository); } }
@@ -2400,10 +2407,15 @@ static int http_request_reauth(const char *url,/* Handle rate limiting with retry logic */ int retry_attempt = http_max_retries - rate_limit_retries +1 ; +trace2_data_intmax ("http" , the_repository,"http/429-retry-attempt" , + retry_attempt); +if (rate_limit_retries <=0 ) {/* Retries are disabled or exhausted */ if (http_max_retries >0 ) {error (_ ("too many rate limit retries, giving up" )); +trace2_data_string ("http" , the_repository, +"http/429-error" ,"retries-exhausted" ); }return HTTP_ERROR ; }
@@ -2418,6 +2430,10 @@ static int http_request_reauth(const char *url,error (_ ("rate limited (HTTP 429) requested %ld second delay, " "exceeds http.maxRetryTime of %ld seconds" ), last_retry_after, http_max_retry_time); +trace2_data_string ("http" , the_repository, +"http/429-error" ,"exceeds-max-retry-time" ); +trace2_data_intmax ("http" , the_repository, +"http/429-requested-delay" , last_retry_after); last_retry_after =-1 ;/* Reset after use */ return HTTP_ERROR ; }
@@ -2429,17 +2445,23 @@ static int http_request_reauth(const char *url,/* Not configured - exit with error */ error (_ ("rate limited (HTTP 429) and no Retry-After header provided. " "Configure http.retryAfter or set GIT_HTTP_RETRY_AFTER." )); +trace2_data_string ("http" , the_repository, +"http/429-error" ,"no-retry-after-config" );return HTTP_ERROR ; } -/* Check if configured default exceeds maximum allowed */ -if (http_retry_after > http_max_retry_time) { -error (_ ("configured http.retryAfter (%ld seconds) exceeds " -"http.maxRetryTime (%ld seconds)" ), - http_retry_after, http_max_retry_time); -return HTTP_ERROR ; - } -/* Use configured default retry-after value */ -sleep_for_retry (http_retry_after); +/* Check if configured default exceeds maximum allowed */ +if (http_retry_after > http_max_retry_time) { +error (_ ("configured http.retryAfter (%ld seconds) exceeds " +"http.maxRetryTime (%ld seconds)" ), + http_retry_after, http_max_retry_time); +trace2_data_string ("http" , the_repository, +"http/429-error" ,"config-exceeds-max-retry-time" ); +return HTTP_ERROR ; + } +/* Use configured default retry-after value */ +trace2_data_string ("http" , the_repository, +"http/429-retry-source" ,"config-default" ); +sleep_for_retry (http_retry_after); } }else if (ret ==HTTP_REAUTH ) {credential_fill (the_repository, &http_auth,1 );
Signature · 1 line
gitgitgadget
Quoted 10 lines · Vaidas Pilkauskas via GitGitGadget
Retry behavior is controlled by three new configuration options:
* http.maxRetries: Maximum number of retry attempts (default: 0, meaning retries are disabled by default). Users must explicitly opt-in to retry behavior.
* http.retryAfter: Default delay in seconds when the server doesn't provide a Retry-After header (default: -1, meaning fail if no header is provided). This serves as a fallback mechanism.
Quoted 8 lines
* http.maxRetryTime: Maximum delay in seconds for a single retry (default: 300). If the server requests a delay exceeding this limit, Git fails immediately rather than waiting. This prevents indefinite blocking on unreasonable server requests.
All three options can be overridden via environment variables:
GIT_HTTP_MAX_RETRIES,GIT_HTTP_RETRY_AFTER, andGIT_HTTP_MAX_RETRY_TIME.
This is great information, and I am glad that it is written down in http.adoc so that it shows up in git-config(1). I think that it's fine to omit this level of detail from the commit message, since it duplicates information from the authoritative source on configuration knobs.
It might be reasonable to say something like:
Retry behavior is controlled by three new configuration options (http.maxRetries, http.retryAfter, and http.maxRetryTime) which are documented in git-config(1).
or something.
http-push.c static int fetch_indices(void) · line 721
Quoted 11 lines
http-push.c +3
-0@@ -716,6 +716,10 @@ static int fetch_indices(void)case HTTP_MISSING_TARGET : ret =0 ;break ; +case HTTP_RATE_LIMITED : +error ("rate limited by '%s', please try again later" , repo->url); + ret =-1 ;
Other strings in this file aren't marked for translation, but I think we can/should mark this one like so:
error(_("rate limited by %s ..."), repo->url);Quoted 13 lines
http.c +1
-0@@ -22,6 +22,7 @@#include "object-file.h" #include "odb.h" #include "tempfile.h" +#include "date.h" static struct trace_key trace_curl =TRACE_KEY_INIT (CURL );static int trace_curl_data =1 ;
http.c static char *cached_accept_language; · line 157
Quoted 8 lines
http.c +5
-0@@ -149,6 +150,14 @@ static char *cached_accept_language;static char *http_ssl_backend;static int http_schannel_check_revoke =1 ; + +/* Retry configuration */ +static long http_retry_after =-1 ;/* Default retry-after in seconds when header is missing (-1 means not set, exit with 128) */ +static long http_max_retries =0 ;/* Maximum number of retry attempts (0 means retries are disabled) */ +static long http_max_retry_time =300 ;/* Maximum time to wait for a single retry (default 5 minutes) */
These comments should be OK to drop, the variables indicate what Git configuration they correspond to (e.g., http_retry_after -> http.retryAfter), so git-config(1) is the authoritative source for documentation here.
http.c static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UN · line 271
Quoted 6 lines
http.c +2
-0@@ -257,6 +267,47 @@ static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UN goto exit; } +/* Parse Retry-After header for rate limiting */ +if (skip_iprefix_mem (ptr, size,"retry-after:" , &val, &val_len)) {
Makes sense, though I wonder if we should rename this function, since fwrite_wwwauth is now doing more than just handling WWW-Authenticate headers.
Perhaps we should have a single top-level function that is registered as our CURLOPT_HEADERFUNCTION that dispatches calls to header-specific functions? Otherwise the actual parsing of the Retry-After header looks good to me.
http.c void http_init(struct remote *remote, const char *url, int proactive_auth) · line 1494
Quoted 8 lines
http.c +4
-0@@ -1422,6 +1488,10 @@ void http_init(struct remote *remote, const char *url, int proactive_auth)set_long_from_env (&curl_tcp_keepintvl,"GIT_TCP_KEEPINTVL" );set_long_from_env (&curl_tcp_keepcnt,"GIT_TCP_KEEPCNT" ); +set_long_from_env (&http_retry_after,"GIT_HTTP_RETRY_AFTER" ); +set_long_from_env (&http_max_retries,"GIT_HTTP_MAX_RETRIES" ); +set_long_from_env (&http_max_retry_time,"GIT_HTTP_MAX_RETRY_TIME" ); +
The configuration handling and overrides look good to me.
http.c static int update_url_from_redirect(struct strbuf *base, · line 2341
Quoted 13 lines
http.c +9
-0@@ -2253,19 +2330,36 @@ static int update_url_from_redirect(struct strbuf *base,return 1 ; } +/* +* Sleep for the specified number of seconds before retrying. +*/ +static void sleep_for_retry (long retry_after) +{ +if (retry_after >0 ) { +unsigned int remaining; +warning (_ ("rate limited, waiting %ld seconds before retry" ), retry_after); + remaining =sleep (retry_after);
What should we do if there are other active request slots? It has been a couple of years since I have looked at Git's HTTP code, but I imagine that we should be able to continue processing other requests while waiting for the retry-after period to elapse here.
http.c static int http_request_reauth(const char *url, · line 2399
Quoted 6 lines
http.c +1
-1@@ -2302,7 +2396,54 @@ static int http_request_reauth(const char *url,BUG ("Unknown http_request target" ); } -credential_fill (the_repository, &http_auth,1 ); +if (ret ==HTTP_RATE_LIMITED ) {
Should handling the retry behavior be moved into a separate function? I think that http_request_reauth() might be clearer if it read:
if (ret == HTTP_RATE_LIMITED)
apply_rate_limit(...); /* presumably with a better name */
else
credential_fill(...);, and likewise, should we rename this function as it is no longer just re-authenticating HTTP requests?
Quoted 21 lines
t/t5584-http-429-retry.sh new file +15
-0@@ -0,0 +1,429 @@ +#!/bin/sh + +test_description='test HTTP 429 Too Many Requests retry logic' + +. ./test-lib.sh + +. "$TEST_DIRECTORY"/lib-httpd.sh + +start_httpd + +test_expect_success 'setup test repository' ' + test_commit initial && + git clone --bare . "$HTTPD_DOCUMENT_ROOT_PATH/repo.git" && + git --git-dir="$HTTPD_DOCUMENT_ROOT_PATH/repo.git" config http.receivepack true +'
t/t5584-http-429-retry.sh line 25
Quoted 10 lines
t/t5584-http-429-retry.sh new file +10
-0+ +test_expect_success 'HTTP 429 with retries disabled (maxRetries=0) fails immediately' ' + write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF && + printf "Status: 429 Too Many Requests\r\n" + printf "Retry-After: 1\r\n" + printf "Content-Type: text/plain\r\n" + printf "\r\n" + printf "Rate limited\n" + cat "$1" >/dev/null + EOF
To avoid having to write this script multiple write, you can write it as a separate script in t/lib-httpd and then make sure to list it in prepare_httpd() (from t/lib-httpd.sh).
You can then list it in the apache.conf in the same directory and invoke it however you like. If you need to take in arguments to the script (e.g., to change the Retry-After value), you can use a ScriptAliasMatch instead of a normal ScriptAlias to pass in extra parameters from the URL.
The one-time-script mechanism here will cause the test harness to delete the script after its first (and only) use, which can be useful for some cases but I suspect is not necessary for all of these tests.
t/t5584-http-429-retry.sh line 35
Quoted 10 lines
t/t5584-http-429-retry.sh new file +10
-0+ + # Set maxRetries to 0 (disabled) + test_config http.maxRetries 0 && + test_config http.retryAfter 1 && + + # Should fail immediately without any retry attempt + test_must_fail git ls-remote "$HTTPD_URL/one_time_script/repo.git" 2>err && + + # Verify no retry happened (no "waiting" message in stderr) + ! grep -i "waiting.*retry" err &&
test_grep can be helpful when reading the output of test failures, since it dumps the contents of the file it was searching. Just make sure to write "test_grep !" instead of "! test_grep" (there are a few such instances of the latter that I just wrote patches to clean up).
"! test_grep" isn't *wrong* per-se, but it will pollute the test output with "couldn't find xyz in abc".
I skimmed through the the remainder of the tests since I imagine that they will change substantially after writing the script out explicitly instead of using one-time-script, so I'll hold off on reviewing that portion in more detail until then.
Thanks,
Taylor
Re: [PATCH 2/3] remote-curl: fix memory leak in show_http_message()
In reply to Vaidas Pilkauskas via GitGitGadgetQuoted 13 lines · Vaidas Pilkauskas via GitGitGadget
remote-curl.c +1
-0@@ -371,6 +371,7 @@ static int show_http_message(struct strbuf *type, struct strbuf *charset, struct strbuf *msg) {const char *p, *eol; +struct strbuf msgbuf =STRBUF_INIT ;/* * We only show text/plain parts, as other types are likely
remote-curl.c static int show_http_message(struct strbuf *type, struct strbuf *charset, · line 383
Quoted 5 lines
remote-curl.c +2
-0@@ -378,19 +379,24 @@ static int show_http_message(struct strbuf *type, struct strbuf *charset,*/ if (strcmp (type->buf,"text/plain" ))return -1 ; + +strbuf_addbuf (&msgbuf, msg);
Hmm. Looking at the list of show_http_message() callers, it looks like they all follow the pattern of constructing a strbuf "msg", passing it to this function, and then calling die() with some user-friendly message.
I agree that the patch here does address that leak, but I wonder if we should do it in a way that doesn't involve copying the "msg" buffer. One thing we could do is rename 'show_http_message()' to make it clear that it's fatal and then free the re-encoded buffer ourselves (along with the other buffers type and charset), perhaps like so (on top of the previous patch in lieu of this one):
--- 8< ---
remote-curl.c +18 -9
@@ -367,23 +367,25 @@ static void free_discovery(struct discovery *d) } } -static int show_http_message (struct strbuf *type,struct strbuf *charset, -struct strbuf *msg) +static void show_http_message_fatal (struct strbuf *type,struct strbuf *charset, +struct strbuf *msg,const char *fmt, ...) {const char *p, *eol; +va_list ap; +report_fn die_message_routine =get_die_message_routine ();/* * We only show text/plain parts, as other types are likely * to be ugly to look at on the user's terminal. */ if (strcmp (type->buf,"text/plain" )) -return -1 ; + goto out;if (charset->len)strbuf_reencode (msg, charset->buf,get_log_output_encoding ());strbuf_trim (msg);if (!msg->len) -return -1 ; + goto out; p = msg->buf; do {
@@ -391,7 +393,15 @@ static int show_http_message(struct strbuf *type, struct strbuf *charset,fprintf (stderr,"remote: %.*s\n" , (int )(eol - p), p); p = eol +1 ; }while (*eol); -return 0 ; + +out: +strbuf_release (type); +strbuf_release (charset); +strbuf_release (msg); + +va_start (ap, fmt); +die_message_routine (fmt, ap); +va_end (ap); } static int get_protocol_http_header(enum protocol_version version,
@@ -518,25 +528,27 @@ static struct discovery *discover_refs(const char *service, int for_push)case HTTP_OK :break ;case HTTP_MISSING_TARGET : -show_http_message (&type, &charset, &buffer); -die (_ ("repository '%s' not found" ), -transport_anonymize_url (url.buf)); +show_http_message_fatal (&type, &charset, &buffer, +_ ("repository '%s' not found" ), +transport_anonymize_url (url.buf)); --- >8 ---
(...and so on for the remaining cases).
Thanks,
Taylor
Quoted 6 lines · Taylor Blau
Quoted 6 lines
http.c +6
-0+/* +* Sleep for the specified number of seconds before retrying. +*/ +static void sleep_for_retry (long retry_after) +{ +if (retry_after >0 ) {
Quoted 8 lines
http.c static int update_url_from_redirect(struct strbuf *base, · line 2341
Quoted 3 lines
http.c +3
-0+unsigned int remaining; +warning (_ ("rate limited, waiting %ld seconds before retry" ), retry_after); + remaining =sleep (retry_after);What should we do if there are other active request slots? It has been a couple of years since I have looked at Git's HTTP code, but I imagine that we should be able to continue processing other requests while waiting for the retry-after period to elapse here.
This is a very good catch - I'll rewrite this to a non-blocking wait.
Thanks for the review, Taylor, I'll work to address this and other comments in the next version of the patch.
Quoted 213 lines · Taylor Blau
Quoted 10 lines · Vaidas Pilkauskas via GitGitGadget
Retry behavior is controlled by three new configuration options:
* http.maxRetries: Maximum number of retry attempts (default: 0, meaning retries are disabled by default). Users must explicitly opt-in to retry behavior.
* http.retryAfter: Default delay in seconds when the server doesn't provide a Retry-After header (default: -1, meaning fail if no header is provided). This serves as a fallback mechanism.
Quoted 8 lines
* http.maxRetryTime: Maximum delay in seconds for a single retry (default: 300). If the server requests a delay exceeding this limit, Git fails immediately rather than waiting. This prevents indefinite blocking on unreasonable server requests.
All three options can be overridden via environment variables:
GIT_HTTP_MAX_RETRIES,GIT_HTTP_RETRY_AFTER, andGIT_HTTP_MAX_RETRY_TIME.This is great information, and I am glad that it is written down in http.adoc so that it shows up in git-config(1). I think that it's fine to omit this level of detail from the commit message, since it duplicates information from the authoritative source on configuration knobs.
It might be reasonable to say something like:
Retry behavior is controlled by three new configuration options (http.maxRetries, http.retryAfter, and http.maxRetryTime) which are documented in git-config(1).
or something.
http-push.c static int fetch_indices(void) · line 721
Quoted 11 lines
http-push.c +3
-0@@ -716,6 +716,10 @@ static int fetch_indices(void)case HTTP_MISSING_TARGET : ret =0 ;break ; +case HTTP_RATE_LIMITED : +error ("rate limited by '%s', please try again later" , repo->url); + ret =-1 ;Other strings in this file aren't marked for translation, but I think we can/should mark this one like so:
error(_("rate limited by %s ..."), repo->url);Quoted 13 lines
http.c +1
-0@@ -22,6 +22,7 @@#include "object-file.h" #include "odb.h" #include "tempfile.h" +#include "date.h" static struct trace_key trace_curl =TRACE_KEY_INIT (CURL );static int trace_curl_data =1 ;http.c static char *cached_accept_language; · line 157
Quoted 8 lines
http.c +5
-0@@ -149,6 +150,14 @@ static char *cached_accept_language;static char *http_ssl_backend;static int http_schannel_check_revoke =1 ; + +/* Retry configuration */ +static long http_retry_after =-1 ;/* Default retry-after in seconds when header is missing (-1 means not set, exit with 128) */ +static long http_max_retries =0 ;/* Maximum number of retry attempts (0 means retries are disabled) */ +static long http_max_retry_time =300 ;/* Maximum time to wait for a single retry (default 5 minutes) */ These comments should be OK to drop, the variables indicate what Git configuration they correspond to (e.g.,
http_retry_after-> http.retryAfter), so git-config(1) is the authoritative source for documentation here.http.c static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UN · line 271
Quoted 6 lines
http.c +2
-0@@ -257,6 +267,47 @@ static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UN goto exit; } +/* Parse Retry-After header for rate limiting */ +if (skip_iprefix_mem (ptr, size,"retry-after:" , &val, &val_len)) {Makes sense, though I wonder if we should rename this function, since
fwrite_wwwauthis now doing more than just handling WWW-Authenticate headers.Perhaps we should have a single top-level function that is registered as our
CURLOPT_HEADERFUNCTIONthat dispatches calls to header-specific functions? Otherwise the actual parsing of the Retry-After header looks good to me.http.c void http_init(struct remote *remote, const char *url, int proactive_auth) · line 1494
Quoted 8 lines
http.c +4
-0@@ -1422,6 +1488,10 @@ void http_init(struct remote *remote, const char *url, int proactive_auth)set_long_from_env (&curl_tcp_keepintvl,"GIT_TCP_KEEPINTVL" );set_long_from_env (&curl_tcp_keepcnt,"GIT_TCP_KEEPCNT" ); +set_long_from_env (&http_retry_after,"GIT_HTTP_RETRY_AFTER" ); +set_long_from_env (&http_max_retries,"GIT_HTTP_MAX_RETRIES" ); +set_long_from_env (&http_max_retry_time,"GIT_HTTP_MAX_RETRY_TIME" ); +The configuration handling and overrides look good to me.
http.c static int update_url_from_redirect(struct strbuf *base, · line 2341
Quoted 13 lines
http.c +9
-0@@ -2253,19 +2330,36 @@ static int update_url_from_redirect(struct strbuf *base,return 1 ; } +/* +* Sleep for the specified number of seconds before retrying. +*/ +static void sleep_for_retry (long retry_after) +{ +if (retry_after >0 ) { +unsigned int remaining; +warning (_ ("rate limited, waiting %ld seconds before retry" ), retry_after); + remaining =sleep (retry_after);What should we do if there are other active request slots? It has been a couple of years since I have looked at Git's HTTP code, but I imagine that we should be able to continue processing other requests while waiting for the retry-after period to elapse here.
http.c static int http_request_reauth(const char *url, · line 2399
Quoted 6 lines
http.c +1
-1@@ -2302,7 +2396,54 @@ static int http_request_reauth(const char *url,BUG ("Unknown http_request target" ); } -credential_fill (the_repository, &http_auth,1 ); +if (ret ==HTTP_RATE_LIMITED ) {Should handling the retry behavior be moved into a separate function? I think that
http_request_reauth()might be clearer if it read:if (ret == HTTP_RATE_LIMITED) apply_rate_limit(...); /* presumably with a better name */ else credential_fill(...);, and likewise, should we rename this function as it is no longer just re-authenticating HTTP requests?
Quoted 21 lines
t/t5584-http-429-retry.sh new file +15
-0@@ -0,0 +1,429 @@ +#!/bin/sh + +test_description='test HTTP 429 Too Many Requests retry logic' + +. ./test-lib.sh + +. "$TEST_DIRECTORY"/lib-httpd.sh + +start_httpd + +test_expect_success 'setup test repository' ' + test_commit initial && + git clone --bare . "$HTTPD_DOCUMENT_ROOT_PATH/repo.git" && + git --git-dir="$HTTPD_DOCUMENT_ROOT_PATH/repo.git" config http.receivepack true +'t/t5584-http-429-retry.sh line 25
Quoted 10 lines
t/t5584-http-429-retry.sh new file +10
-0+ +test_expect_success 'HTTP 429 with retries disabled (maxRetries=0) fails immediately' ' + write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF && + printf "Status: 429 Too Many Requests\r\n" + printf "Retry-After: 1\r\n" + printf "Content-Type: text/plain\r\n" + printf "\r\n" + printf "Rate limited\n" + cat "$1" >/dev/null + EOFTo avoid having to write this script multiple write, you can write it as a separate script in t/lib-httpd and then make sure to list it in
prepare_httpd()(fromt/lib-httpd.sh).You can then list it in the apache.conf in the same directory and invoke it however you like. If you need to take in arguments to the script (e.g., to change the Retry-After value), you can use a ScriptAliasMatch instead of a normal ScriptAlias to pass in extra parameters from the URL.
The one-time-script mechanism here will cause the test harness to delete the script after its first (and only) use, which can be useful for some cases but I suspect is not necessary for all of these tests.
t/t5584-http-429-retry.sh line 35
Quoted 10 lines
t/t5584-http-429-retry.sh new file +10
-0+ + # Set maxRetries to 0 (disabled) + test_config http.maxRetries 0 && + test_config http.retryAfter 1 && + + # Should fail immediately without any retry attempt + test_must_fail git ls-remote "$HTTPD_URL/one_time_script/repo.git" 2>err && + + # Verify no retry happened (no "waiting" message in stderr) + ! grep -i "waiting.*retry" err &&
test_grepcan be helpful when reading the output of test failures, since it dumps the contents of the file it was searching. Just make sure to write "test_grep!" instead of "!test_grep" (there are a few such instances of the latter that I just wrote patches to clean up)."!
test_grep" isn't *wrong* per-se, but it will pollute the test output with "couldn't find xyz in abc".I skimmed through the the remainder of the tests since I imagine that they will change substantially after writing the script out explicitly instead of using one-time-script, so I'll hold off on reviewing that portion in more detail until then.
Thanks,
Taylor
v2 cover http: add support for HTTP 429 rate limit retries
In reply to Vaidas Pilkauskas via GitGitGadgetChanges since v1:
* removed configuration options from commit message
* marked "rate limited by %s ..." message for translation in http-push.c
* dropped redundant comments from retry configuration variables
* renamed "fwrite_wwwauth" to "fwrite_headers". Due to complexity related with header continuation handling, I've decide not to split into header specific functions.
* rewritten the sleep logic into non-blocking handling, so that the rest of the slots can be processed at the time retry delay is requested
* renamed "http_request_reauth" to "http_request_recoverable" to better reflect that it does not only reauth , but also other recoverable handling
* updated test setup code to use setup script to reduce repetition in the test code
* updated to use test_grep instead of grep
* dropped memory leak fix patch in afvor to rename of the "show_http_message" to "show_http_message_fatal".
* Adjusted alloc size in "strbuf_reencode" for "strbuf_attach" call, which seems to solve the leak problem
The implementation includes:
Patch 1: Core HTTP 429 retry logic with support for RFC-compliant Retry-After headers (both delay-seconds and HTTP-date formats), comprehensive configuration options, and fail-fast behavior for excessive delays. Includes extensive test coverage.
Patch 2: Adds trace2 instrumentation to enable monitoring and debugging of retry operations in production environments.
Vaidas Pilkauskas (2):
http: add support for HTTP 429 rate limit retries
http: add trace2 logging for retry operations
Documentation/config/http.adoc | 24 +++ http-push.c | 8 + http-walker.c | 5 + http.c | 321 +++++++++++++++++++++++++++++---- http.h | 4 + remote-curl.c | 49 +++-- strbuf.c | 2 +- t/lib-httpd.sh | 1 + t/lib-httpd/apache.conf | 8 + t/lib-httpd/http-429.sh | 98 ++++++++++ t/meson.build | 1 + t/t5584-http-429-retry.sh | 286 +++++++++++++++++++++++++++++ 12 files changed, 750 insertions(+), 57 deletions(-) create mode 100644 t/lib-httpd/http-429.sh create mode 100755 t/t5584-http-429-retry.sh
base-commit: c4a0c8845e2426375ad257b6c221a3a7d92ecfda
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-2008%2Fvaidas-shopify%2Fretry-after-v2
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-2008/vaidas-shopify/retry-after-v2
Pull-Request: https://github.com/gitgitgadget/git/pull/2008
Range-diff vs v1:
1: ae0087cd1c ! 1: d80ce07703 http: add support for HTTP 429 rate limit retries
@@ Commit message
both delay-seconds (integer) and HTTP-date (RFC 2822) formats. If a
past date is provided, Git retries immediately without waiting.
- Retry behavior is controlled by three new configuration options:
-
- * http.maxRetries: Maximum number of retry attempts (default: 0,
- meaning retries are disabled by default). Users must explicitly
- opt-in to retry behavior.
-
- * http.retryAfter: Default delay in seconds when the server doesn't
- provide a Retry-After header (default: -1, meaning fail if no
- header is provided). This serves as a fallback mechanism.
-
- * http.maxRetryTime: Maximum delay in seconds for a single retry
- (default: 300). If the server requests a delay exceeding this
- limit, Git fails immediately rather than waiting. This prevents
- indefinite blocking on unreasonable server requests.
-
- All three options can be overridden via environment variables:
- GIT_HTTP_MAX_RETRIES, GIT_HTTP_RETRY_AFTER, and
- GIT_HTTP_MAX_RETRY_TIME.
+ Retry behavior is controlled by three new configuration options
+ (http.maxRetries, http.retryAfter, and http.maxRetryTime) which are
+ documented in git-config(1).
The retry logic implements a fail-fast approach: if any delay
(whether from server header or configuration) exceeds maxRetryTime,
@@ http-push.c: static int fetch_indices(void)
ret = 0;
break;
+ case HTTP_RATE_LIMITED:
-+ error("rate limited by '%s', please try again later", repo->url);
++ error(_("rate limited by '%s', please try again later"), repo->url);
+ ret = -1;
+ break;
default:
@@ http-push.c: static int remote_exists(const char *path)
ret = 0;
break;
+ case HTTP_RATE_LIMITED:
-+ error("rate limited by '%s', please try again later", url);
++ error(_("rate limited by '%s', please try again later"), url);
+ ret = -1;
+ break;
case HTTP_ERROR:
@@ http.c: static char *cached_accept_language;
static int http_schannel_check_revoke = 1;
+
-+/* Retry configuration */
-+static long http_retry_after = -1; /* Default retry-after in seconds when header is missing (-1 means not set, exit with 128) */
-+static long http_max_retries = 0; /* Maximum number of retry attempts (0 means retries are disabled) */
-+static long http_max_retry_time = 300; /* Maximum time to wait for a single retry (default 5 minutes) */
++static long http_retry_after = -1;
++static long http_max_retries = 0;
++static long http_max_retry_time = 300;
+
-+/* Store retry_after value from 429 responses for retry logic (-1 = not set, 0 = retry immediately, >0 = delay in seconds) */
-+static long last_retry_after = -1;
/*
* With the backend being set to `schannel`, setting sslCAinfo would override
* the Certificate Store in cURL v7.60.0 and later, which is not what we want
@@ http.c: static inline int is_hdr_continuation(const char *ptr, const size_t size
}
-static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UNUSED)
-+static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p)
++static size_t fwrite_headers(char *ptr, size_t eltsize, size_t nmemb, void *p)
{
size_t size = eltsize * nmemb;
struct strvec *values = &http_auth.wwwauth_headers;
@@ http.c: void http_init(struct remote *remote, const char *url, int proactive_aut
curl_default = get_curl_handle();
}
+@@ http.c: struct active_request_slot *get_active_slot(void)
+ slot->finished = NULL;
+ slot->callback_data = NULL;
+ slot->callback_func = NULL;
++ slot->retry_delay_seconds = -1;
++ memset(&slot->retry_delay_start, 0, sizeof(slot->retry_delay_start));
+
+ if (curl_cookie_file && !strcmp(curl_cookie_file, "-")) {
+ warning(_("refusing to read cookies from http.cookiefile '-'"));
+@@ http.c: void run_active_slot(struct active_request_slot *slot)
+ fd_set excfds;
+ int max_fd;
+ struct timeval select_timeout;
++ long curl_timeout;
++ struct timeval start_time = {0}, current_time, elapsed_time = {0};
++ long remaining_seconds;
+ int finished = 0;
++ int slot_not_started = (slot->finished == NULL);
++ int waiting_for_delay = (slot->retry_delay_seconds > 0);
++
++ if (waiting_for_delay) {
++ warning(_("rate limited, waiting %ld seconds before retry"), slot->retry_delay_seconds);
++ start_time = slot->retry_delay_start;
++ }
+
+ slot->finished = &finished;
+- while (!finished) {
++ while (waiting_for_delay || !finished) {
++ if (waiting_for_delay) {
++ gettimeofday(¤t_time, NULL);
++ elapsed_time.tv_sec = current_time.tv_sec - start_time.tv_sec;
++ elapsed_time.tv_usec = current_time.tv_usec - start_time.tv_usec;
++ if (elapsed_time.tv_usec < 0) {
++ elapsed_time.tv_sec--;
++ elapsed_time.tv_usec += 1000000;
++ }
++
++ if (elapsed_time.tv_sec >= slot->retry_delay_seconds) {
++ slot->retry_delay_seconds = -1;
++ waiting_for_delay = 0;
++
++ if (slot_not_started)
++ return;
++ }
++ }
++
+ step_active_slots();
+
+- if (slot->in_use) {
+- long curl_timeout;
+- curl_multi_timeout(curlm, &curl_timeout);
+- if (curl_timeout == 0) {
++ if (!waiting_for_delay && !slot->in_use)
++ continue;
++
++ curl_multi_timeout(curlm, &curl_timeout);
++ if (curl_timeout == 0) {
++ if (!waiting_for_delay)
+ continue;
+- } else if (curl_timeout == -1) {
+- select_timeout.tv_sec = 0;
+- select_timeout.tv_usec = 50000;
++ select_timeout.tv_sec = 0;
++ select_timeout.tv_usec = 50000; /* 50ms */
++ } else if (curl_timeout == -1) {
++ select_timeout.tv_sec = 0;
++ select_timeout.tv_usec = 50000;
++ } else {
++ long curl_timeout_sec = curl_timeout / 1000;
++ long curl_timeout_usec = (curl_timeout % 1000) * 1000;
++
++ if (waiting_for_delay) {
++ remaining_seconds = slot->retry_delay_seconds - elapsed_time.tv_sec;
++ if (curl_timeout_sec < remaining_seconds) {
++ select_timeout.tv_sec = curl_timeout_sec;
++ select_timeout.tv_usec = curl_timeout_usec;
++ } else {
++ select_timeout.tv_sec = remaining_seconds;
++ select_timeout.tv_usec = 0;
++ }
+ } else {
+- select_timeout.tv_sec = curl_timeout / 1000;
+- select_timeout.tv_usec = (curl_timeout % 1000) * 1000;
++ select_timeout.tv_sec = curl_timeout_sec;
++ select_timeout.tv_usec = curl_timeout_usec;
+ }
++ }
+
+- max_fd = -1;
+- FD_ZERO(&readfds);
+- FD_ZERO(&writefds);
+- FD_ZERO(&excfds);
+- curl_multi_fdset(curlm, &readfds, &writefds, &excfds, &max_fd);
++ max_fd = -1;
++ FD_ZERO(&readfds);
++ FD_ZERO(&writefds);
++ FD_ZERO(&excfds);
++ curl_multi_fdset(curlm, &readfds, &writefds, &excfds, &max_fd);
+
+- /*
+- * It can happen that curl_multi_timeout returns a pathologically
+- * long timeout when curl_multi_fdset returns no file descriptors
+- * to read. See commit message for more details.
+- */
+- if (max_fd < 0 &&
+- (select_timeout.tv_sec > 0 ||
+- select_timeout.tv_usec > 50000)) {
+- select_timeout.tv_sec = 0;
+- select_timeout.tv_usec = 50000;
+- }
++ /*
++ * It can happen that curl_multi_timeout returns a pathologically
++ * long timeout when curl_multi_fdset returns no file descriptors
++ * to read. See commit message for more details.
++ */
++ if (max_fd < 0 &&
++ (select_timeout.tv_sec > 0 ||
++ select_timeout.tv_usec > 50000)) {
++ select_timeout.tv_sec = 0;
++ select_timeout.tv_usec = 50000;
++ }
+
+- select(max_fd+1, &readfds, &writefds, &excfds, &select_timeout);
++ /*
++ * If curl_multi_fdset returns no file descriptors but we have
++ * a timeout, still use select() to wait for the timeout period.
++ */
++ if (max_fd < 0) {
++ /* No file descriptors, just wait for timeout */
++ select(0, NULL, NULL, NULL, &select_timeout);
++ } else {
++ select(max_fd + 1, &readfds, &writefds, &excfds, &select_timeout);
+ }
+ }
+
@@ http.c: static int handle_curl_result(struct slot_results *results)
}
return HTTP_REAUTH;
}
+ } else if (results->http_code == 429) {
-+ /* Store the retry_after value for use in retry logic */
-+ last_retry_after = results->retry_after;
+ return HTTP_RATE_LIMITED;
} else {
if (results->http_connectcode == 407)
@@ http.c: int run_one_slot(struct active_request_slot *slot,
slot->results = results;
+ /* Initialize retry_after to -1 (not set) */
+ results->retry_after = -1;
++
++ /* If there's a retry delay, wait for it before starting the slot */
++ if (slot->retry_delay_seconds > 0) {
++ run_active_slot(slot);
++ }
++
if (!start_active_slot(slot)) {
xsnprintf(curl_errorstr, sizeof(curl_errorstr),
"failed to start HTTP request");
+@@ http.c: static void http_opt_request_remainder(CURL *curl, off_t pos)
+ #define HTTP_REQUEST_STRBUF 0
+ #define HTTP_REQUEST_FILE 1
+
++static void sleep_for_retry(struct active_request_slot *slot, long retry_after);
++
+ static int http_request(const char *url,
+ void *result, int target,
+- const struct http_get_options *options)
++ const struct http_get_options *options,
++ long *retry_after_out,
++ long retry_delay)
+ {
+ struct active_request_slot *slot;
+ struct slot_results results;
@@ http.c: static int http_request(const char *url,
+ int ret;
+
+ slot = get_active_slot();
++ /* Mark slot for delay if retry delay is provided */
++ if (retry_delay > 0) {
++ sleep_for_retry(slot, retry_delay);
++ }
+ curl_easy_setopt(slot->curl, CURLOPT_HTTPGET, 1L);
+
+ if (!result) {
+@@ http.c: static int http_request(const char *url,
+ fwrite_buffer);
}
- curl_easy_setopt(slot->curl, CURLOPT_HEADERFUNCTION, fwrite_wwwauth);
+- curl_easy_setopt(slot->curl, CURLOPT_HEADERFUNCTION, fwrite_wwwauth);
++ curl_easy_setopt(slot->curl, CURLOPT_HEADERFUNCTION, fwrite_headers);
+ curl_easy_setopt(slot->curl, CURLOPT_HEADERDATA, slot);
accept_language = http_get_accept_language_header();
+@@ http.c: static int http_request(const char *url,
+
+ ret = run_one_slot(slot, &results);
+
++ /* Store retry_after from slot results if output parameter provided */
++ if (retry_after_out)
++ *retry_after_out = results.retry_after;
++
+ if (options && options->content_type) {
+ struct strbuf raw = STRBUF_INIT;
+ curlinfo_strbuf(slot->curl, CURLINFO_CONTENT_TYPE, &raw);
@@ http.c: static int update_url_from_redirect(struct strbuf *base,
return 1;
}
+-static int http_request_reauth(const char *url,
+/*
-+ * Sleep for the specified number of seconds before retrying.
++ * Mark slot to be delayed for retry. The actual delay will be handled
++ * in run_active_slot when the slot is executed.
+ */
-+static void sleep_for_retry(long retry_after)
++static void sleep_for_retry(struct active_request_slot *slot, long retry_after)
+{
-+ if (retry_after > 0) {
-+ unsigned int remaining;
++ if (retry_after > 0 && slot) {
+ warning(_("rate limited, waiting %ld seconds before retry"), retry_after);
-+ remaining = sleep(retry_after);
-+ while (remaining > 0) {
-+ /* Sleep was interrupted, continue sleeping */
-+ remaining = sleep(remaining);
++ slot->retry_delay_seconds = retry_after;
++ gettimeofday(&slot->retry_delay_start, NULL);
++ }
++}
++
++/*
++ * Handle rate limiting retry logic for HTTP 429 responses.
++ * Uses slot-specific retry_after value to support concurrent slots.
++ * Returns a negative value if retries are exhausted or configuration is invalid,
++ * otherwise returns the delay value (>= 0) to indicate the retry should proceed.
++ */
++static long handle_rate_limit_retry(int *rate_limit_retries, long slot_retry_after)
++{
++ int retry_attempt = http_max_retries - *rate_limit_retries + 1;
++ if (*rate_limit_retries <= 0) {
++ /* Retries are disabled or exhausted */
++ if (http_max_retries > 0) {
++ error(_("too many rate limit retries, giving up"));
++ }
++ return -1;
++ }
++
++ /* Decrement retries counter */
++ (*rate_limit_retries)--;
++
++ /* Use the slot-specific retry_after value or configured default */
++ if (slot_retry_after >= 0) {
++ /* Check if retry delay exceeds maximum allowed */
++ if (slot_retry_after > http_max_retry_time) {
++ error(_("rate limited (HTTP 429) requested %ld second delay, "
++ "exceeds http.maxRetryTime of %ld seconds"),
++ slot_retry_after, http_max_retry_time);
++ return -1;
++ }
++ return slot_retry_after;
++ } else {
++ /* No Retry-After header provided */
++ if (http_retry_after < 0) {
++ /* Not configured - exit with error */
++ error(_("rate limited (HTTP 429) and no Retry-After header provided. "
++ "Configure http.retryAfter or set GIT_HTTP_RETRY_AFTER."));
++ return -1;
+ }
++ /* Check if configured default exceeds maximum allowed */
++ if (http_retry_after > http_max_retry_time) {
++ error(_("configured http.retryAfter (%ld seconds) exceeds "
++ "http.maxRetryTime (%ld seconds)"),
++ http_retry_after, http_max_retry_time);
++ return -1;
++ }
++
++ return http_retry_after;
+ }
+}
+
- static int http_request_reauth(const char *url,
++static int http_request_recoverable(const char *url,
void *result, int target,
struct http_get_options *options)
{
int i = 3;
int ret;
+ int rate_limit_retries = http_max_retries;
++ long slot_retry_after = -1; /* Per-slot retry_after value */
if (always_auth_proactively())
credential_fill(the_repository, &http_auth, 1);
- ret = http_request(url, result, target, options);
+- ret = http_request(url, result, target, options);
++ ret = http_request(url, result, target, options, &slot_retry_after, -1);
- if (ret != HTTP_OK && ret != HTTP_REAUTH)
+ if (ret != HTTP_OK && ret != HTTP_REAUTH && ret != HTTP_RATE_LIMITED)
return ret;
++ /* If retries are disabled and we got a 429, fail immediately */
++ if (ret == HTTP_RATE_LIMITED && http_max_retries == 0)
++ return HTTP_ERROR;
++
if (options && options->effective_url && options->base_url) {
+ if (update_url_from_redirect(options->base_url,
+ url, options->effective_url)) {
@@ http.c: static int http_request_reauth(const char *url,
}
}
- while (ret == HTTP_REAUTH && --i) {
+ while ((ret == HTTP_REAUTH || ret == HTTP_RATE_LIMITED) && --i) {
++ long retry_delay = -1;
/*
* The previous request may have put cruft into our output stream; we
* should clear it out before making our next request.
@@ http.c: static int http_request_reauth(const char *url,
+ default:
BUG("Unknown http_request target");
}
-
-- credential_fill(the_repository, &http_auth, 1);
+ if (ret == HTTP_RATE_LIMITED) {
-+ /* Handle rate limiting with retry logic */
-+ int retry_attempt = http_max_retries - rate_limit_retries + 1;
-+
-+ if (rate_limit_retries <= 0) {
-+ /* Retries are disabled or exhausted */
-+ if (http_max_retries > 0) {
-+ error(_("too many rate limit retries, giving up"));
-+ }
++ retry_delay = handle_rate_limit_retry(&rate_limit_retries, slot_retry_after);
++ if (retry_delay < 0)
+ return HTTP_ERROR;
-+ }
-+
-+ /* Decrement retries counter */
-+ rate_limit_retries--;
-+
-+ /* Use the stored retry_after value or configured default */
-+ if (last_retry_after >= 0) {
-+ /* Check if retry delay exceeds maximum allowed */
-+ if (last_retry_after > http_max_retry_time) {
-+ error(_("rate limited (HTTP 429) requested %ld second delay, "
-+ "exceeds http.maxRetryTime of %ld seconds"),
-+ last_retry_after, http_max_retry_time);
-+ last_retry_after = -1; /* Reset after use */
-+ return HTTP_ERROR;
-+ }
-+ sleep_for_retry(last_retry_after);
-+ last_retry_after = -1; /* Reset after use */
-+ } else {
-+ /* No Retry-After header provided */
-+ if (http_retry_after < 0) {
-+ /* Not configured - exit with error */
-+ error(_("rate limited (HTTP 429) and no Retry-After header provided. "
-+ "Configure http.retryAfter or set GIT_HTTP_RETRY_AFTER."));
-+ return HTTP_ERROR;
-+ }
-+ /* Check if configured default exceeds maximum allowed */
-+ if (http_retry_after > http_max_retry_time) {
-+ error(_("configured http.retryAfter (%ld seconds) exceeds "
-+ "http.maxRetryTime (%ld seconds)"),
-+ http_retry_after, http_max_retry_time);
-+ return HTTP_ERROR;
-+ }
-+ /* Use configured default retry-after value */
-+ sleep_for_retry(http_retry_after);
-+ }
++ slot_retry_after = -1; /* Reset after use */
+ } else if (ret == HTTP_REAUTH) {
+ credential_fill(the_repository, &http_auth, 1);
+ }
- ret = http_request(url, result, target, options);
+- credential_fill(the_repository, &http_auth, 1);
+-
+- ret = http_request(url, result, target, options);
++ ret = http_request(url, result, target, options, &slot_retry_after, retry_delay);
}
+ return ret;
+ }
+@@ http.c: int http_get_strbuf(const char *url,
+ struct strbuf *result,
+ struct http_get_options *options)
+ {
+- return http_request_reauth(url, result, HTTP_REQUEST_STRBUF, options);
++ return http_request_recoverable(url, result, HTTP_REQUEST_STRBUF, options);
+ }
+
+ /*
+@@ http.c: int http_get_file(const char *url, const char *filename,
+ goto cleanup;
+ }
+
+- ret = http_request_reauth(url, result, HTTP_REQUEST_FILE, options);
++ ret = http_request_recoverable(url, result, HTTP_REQUEST_FILE, options);
+ fclose(result);
+
+ if (ret == HTTP_OK && finalize_object_file(the_repository, tmpfile.buf, filename))
## http.h ##
@@ http.h: struct slot_results {
@@ http.h: struct slot_results {
};
struct active_request_slot {
+@@ http.h: struct active_request_slot {
+ void *callback_data;
+ void (*callback_func)(void *data);
+ struct active_request_slot *next;
++ long retry_delay_seconds;
++ struct timeval retry_delay_start;
+ };
+
+ struct buffer {
@@ http.h: struct http_get_options {
#define HTTP_REAUTH 4
#define HTTP_NOAUTH 5
@@ http.h: struct http_get_options {
* Requests a URL and stores the result in a strbuf.
## remote-curl.c ##
+@@ remote-curl.c: static void free_discovery(struct discovery *d)
+ }
+ }
+
+-static int show_http_message(struct strbuf *type, struct strbuf *charset,
+- struct strbuf *msg)
++static NORETURN void show_http_message_fatal(struct strbuf *type, struct strbuf *charset,
++ struct strbuf *msg, const char *fmt, ...)
+ {
+ const char *p, *eol;
++ va_list ap;
++ report_fn die_message_routine = get_die_message_routine();
+
+ /*
+ * We only show text/plain parts, as other types are likely
+ * to be ugly to look at on the user's terminal.
+ */
+ if (strcmp(type->buf, "text/plain"))
+- return -1;
++ goto out;
+ if (charset->len)
+ strbuf_reencode(msg, charset->buf, get_log_output_encoding());
+
+ strbuf_trim(msg);
+ if (!msg->len)
+- return -1;
++ goto out;
+
+ p = msg->buf;
+ do {
+@@ remote-curl.c: static int show_http_message(struct strbuf *type, struct strbuf *charset,
+ fprintf(stderr, "remote: %.*s\n", (int)(eol - p), p);
+ p = eol + 1;
+ } while(*eol);
+- return 0;
++
++out:
++ strbuf_release(type);
++ strbuf_release(charset);
++ strbuf_release(msg);
++
++ va_start(ap, fmt);
++ die_message_routine(fmt, ap);
++ va_end(ap);
++ exit(128);
+ }
+
+ static int get_protocol_http_header(enum protocol_version version,
@@ remote-curl.c: static struct discovery *discover_refs(const char *service, int for_push)
- show_http_message(&type, &charset, &buffer);
- die(_("unable to access '%s' with http.pinnedPubkey configuration: %s"),
- transport_anonymize_url(url.buf), curl_errorstr);
+ case HTTP_OK:
+ break;
+ case HTTP_MISSING_TARGET:
+- show_http_message(&type, &charset, &buffer);
+- die(_("repository '%s' not found"),
+- transport_anonymize_url(url.buf));
++ show_http_message_fatal(&type, &charset, &buffer,
++ _("repository '%s' not found"),
++ transport_anonymize_url(url.buf));
+ case HTTP_NOAUTH:
+- show_http_message(&type, &charset, &buffer);
+- die(_("Authentication failed for '%s'"),
+- transport_anonymize_url(url.buf));
++ show_http_message_fatal(&type, &charset, &buffer,
++ _("Authentication failed for '%s'"),
++ transport_anonymize_url(url.buf));
+ case HTTP_NOMATCHPUBLICKEY:
+- show_http_message(&type, &charset, &buffer);
+- die(_("unable to access '%s' with http.pinnedPubkey configuration: %s"),
+- transport_anonymize_url(url.buf), curl_errorstr);
++ show_http_message_fatal(&type, &charset, &buffer,
++ _("unable to access '%s' with http.pinnedPubkey configuration: %s"),
++ transport_anonymize_url(url.buf), curl_errorstr);
+ case HTTP_RATE_LIMITED:
-+ show_http_message(&type, &charset, &buffer);
-+ die(_("rate limited by '%s', please try again later"),
-+ transport_anonymize_url(url.buf));
++ show_http_message_fatal(&type, &charset, &buffer,
++ _("rate limited by '%s', please try again later"),
++ transport_anonymize_url(url.buf));
default:
- show_http_message(&type, &charset, &buffer);
- die(_("unable to access '%s': %s"),
+- show_http_message(&type, &charset, &buffer);
+- die(_("unable to access '%s': %s"),
+- transport_anonymize_url(url.buf), curl_errorstr);
++ show_http_message_fatal(&type, &charset, &buffer,
++ _("unable to access '%s': %s"),
++ transport_anonymize_url(url.buf), curl_errorstr);
+ }
+
+ if (options.verbosity && !starts_with(refs_url.buf, url.buf)) {
+
+ ## strbuf.c ##
+@@ strbuf.c: int strbuf_reencode(struct strbuf *sb, const char *from, const char *to)
+ if (!out)
+ return -1;
+
+- strbuf_attach(sb, out, len, len);
++ strbuf_attach(sb, out, len, len + 1);
+ return 0;
+ }
+
+
+ ## t/lib-httpd.sh ##
+@@ t/lib-httpd.sh: prepare_httpd() {
+ install_script error.sh
+ install_script apply-one-time-script.sh
+ install_script nph-custom-auth.sh
++ install_script http-429.sh
+
+ ln -s "$LIB_HTTPD_MODULE_PATH" "$HTTPD_ROOT_PATH/modules"
+
+
+ ## t/lib-httpd/apache.conf ##
+@@ t/lib-httpd/apache.conf: SetEnv PERL_PATH ${PERL_PATH}
+ SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH}
+ SetEnv GIT_HTTP_EXPORT_ALL
+ </LocationMatch>
++<LocationMatch /http_429/>
++ SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH}
++ SetEnv GIT_HTTP_EXPORT_ALL
++</LocationMatch>
+ <LocationMatch /smart_v0/>
+ SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH}
+ SetEnv GIT_HTTP_EXPORT_ALL
+@@ t/lib-httpd/apache.conf: ScriptAlias /broken_smart/ broken-smart-http.sh/
+ ScriptAlias /error_smart/ error-smart-http.sh/
+ ScriptAlias /error/ error.sh/
+ ScriptAliasMatch /one_time_script/(.*) apply-one-time-script.sh/$1
++ScriptAliasMatch /http_429/(.*) http-429.sh/$1
+ ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1
+ <Directory ${GIT_EXEC_PATH}>
+ Options FollowSymlinks
+@@ t/lib-httpd/apache.conf: ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1
+ <Files apply-one-time-script.sh>
+ Options ExecCGI
+ </Files>
++<Files http-429.sh>
++ Options ExecCGI
++</Files>
+ <Files ${GIT_EXEC_PATH}/git-http-backend>
+ Options ExecCGI
+ </Files>
+
+ ## t/lib-httpd/http-429.sh (new) ##
+@@
++#!/bin/sh
++
++# Script to return HTTP 429 Too Many Requests responses for testing retry logic.
++# Usage: /http_429/<test-context>/<retry-after-value>/<repo-path>
++#
++# The test-context is a unique identifier for each test to isolate state files.
++# The retry-after-value can be:
++# - A number (e.g., "1", "2", "100") - sets Retry-After header to that many seconds
++# - "none" - no Retry-After header
++# - "invalid" - invalid Retry-After format
++# - "permanent" - always return 429 (never succeed)
++# - An HTTP-date string (RFC 2822 format) - sets Retry-After to that date
++#
++# On first call, returns 429. On subsequent calls (after retry), forwards to git-http-backend
++# unless retry-after-value is "permanent".
++
++# Extract test context, retry-after value and repo path from PATH_INFO
++# PATH_INFO format: /<test-context>/<retry-after-value>/<repo-path>
++path_info="${PATH_INFO#/}" # Remove leading slash
++test_context="${path_info%%/*}" # Get first component (test context)
++remaining="${path_info#*/}" # Get rest
++retry_after="${remaining%%/*}" # Get second component (retry-after value)
++repo_path="${remaining#*/}" # Get rest (repo path)
++
++# Extract repository name from repo_path (e.g., "repo.git" from "repo.git/info/refs")
++# The repo name is the first component before any "/"
++repo_name="${repo_path%%/*}"
++
++# Use current directory (HTTPD_ROOT_PATH) for state file
++# Create a safe filename from test_context, retry_after and repo_name
++# This ensures all requests for the same test context share the same state file
++safe_name=$(echo "${test_context}-${retry_after}-${repo_name}" | tr '/' '_' | tr -cd 'a-zA-Z0-9_-')
++state_file="http-429-state-${safe_name}"
++
++# Check if this is the first call (no state file exists)
++if test -f "$state_file"
++then
++ # Already returned 429 once, forward to git-http-backend
++ # Set PATH_INFO to just the repo path (without retry-after value)
++ # Set GIT_PROJECT_ROOT so git-http-backend can find the repository
++ # Use exec to replace this process so git-http-backend gets the updated environment
++ PATH_INFO="/$repo_path"
++ export PATH_INFO
++ # GIT_PROJECT_ROOT points to the document root where repositories are stored
++ # The script runs from HTTPD_ROOT_PATH, and www/ is the document root
++ if test -z "$GIT_PROJECT_ROOT"
++ then
++ # Construct path: current directory (HTTPD_ROOT_PATH) + /www
++ GIT_PROJECT_ROOT="$(pwd)/www"
++ export GIT_PROJECT_ROOT
++ fi
++ exec "$GIT_EXEC_PATH/git-http-backend"
++fi
++
++# Mark that we've returned 429
++touch "$state_file"
++
++# Output HTTP 429 response
++printf "Status: 429 Too Many Requests\r\n"
++
++# Set Retry-After header based on retry_after value
++case "$retry_after" in
++ none)
++ # No Retry-After header
++ ;;
++ invalid)
++ printf "Retry-After: invalid-format-123abc\r\n"
++ ;;
++ permanent)
++ # Always return 429, don't set state file for success
++ rm -f "$state_file"
++ printf "Retry-After: 1\r\n"
++ printf "Content-Type: text/plain\r\n"
++ printf "\r\n"
++ printf "Permanently rate limited\n"
++ exit 0
++ ;;
++ *)
++ # Check if it's a number
++ case "$retry_after" in
++ [0-9]*)
++ # Numeric value
++ printf "Retry-After: %s\r\n" "$retry_after"
++ ;;
++ *)
++ # Assume it's an HTTP-date format (passed as-is, URL decoded)
++ # Apache may URL-encode the path, so decode common URL-encoded characters
++ # %20 = space, %2C = comma, %3A = colon
++ retry_value=$(echo "$retry_after" | sed -e 's/%20/ /g' -e 's/%2C/,/g' -e 's/%3A/:/g')
++ printf "Retry-After: %s\r\n" "$retry_value"
++ ;;
++ esac
++ ;;
++esac
++
++printf "Content-Type: text/plain\r\n"
++printf "\r\n"
++printf "Rate limited\n"
## t/meson.build ##
@@ t/meson.build: integration_tests = [
@@ t/t5584-http-429-retry.sh (new)
+ git --git-dir="$HTTPD_DOCUMENT_ROOT_PATH/repo.git" config http.receivepack true
+'
+
-+test_expect_success 'HTTP 429 with retries disabled (maxRetries=0) fails immediately' '
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF &&
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Retry-After: 1\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Rate limited\n"
-+ cat "$1" >/dev/null
-+ EOF
++# This test suite uses a special HTTP 429 endpoint at /http_429/ that simulates
++# rate limiting. The endpoint format is:
++# /http_429/<test-context>/<retry-after-value>/<repo-path>
++# The http-429.sh script (in t/lib-httpd) returns a 429 response with the
++# specified Retry-After header on the first request for each test context,
++# then forwards subsequent requests to git-http-backend. Each test context
++# is isolated, allowing multiple tests to run independently.
+
++test_expect_success 'HTTP 429 with retries disabled (maxRetries=0) fails immediately' '
+ # Set maxRetries to 0 (disabled)
+ test_config http.maxRetries 0 &&
+ test_config http.retryAfter 1 &&
+
+ # Should fail immediately without any retry attempt
-+ test_must_fail git ls-remote "$HTTPD_URL/one_time_script/repo.git" 2>err &&
++ test_must_fail git ls-remote "$HTTPD_URL/http_429/retries-disabled/1/repo.git" 2>err &&
+
+ # Verify no retry happened (no "waiting" message in stderr)
-+ ! grep -i "waiting.*retry" err &&
-+
-+ # The one-time script will be consumed on first request (not a retry)
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ test_grep ! -i "waiting.*retry" err
+'
+
+test_expect_success 'HTTP 429 permanent should fail after max retries' '
-+ # Install a permanent error script to prove retries are limited
-+ write_script "$HTTPD_ROOT_PATH/http-429-permanent.sh" <<-\EOF &&
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Retry-After: 1\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Permanently rate limited\n"
-+ EOF
-+
+ # Enable retries with a limit
+ test_config http.maxRetries 2 &&
+
+ # Git should retry but eventually fail when 429 persists
-+ test_must_fail git ls-remote "$HTTPD_URL/error/http-429-permanent.sh/repo.git" 2>err
++ test_must_fail git ls-remote "$HTTPD_URL/http_429/permanent-fail/permanent/repo.git" 2>err
+'
+
+test_expect_success 'HTTP 429 with Retry-After is retried and succeeds' '
-+ # Create a one-time script that returns 429 with Retry-After header
-+ # on the first request. Subsequent requests will succeed.
-+ # This contrasts with the permanent 429 above - proving retry works
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF &&
-+ # Return HTTP 429 response instead of git response
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Retry-After: 1\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Rate limited - please retry after 1 second\n"
-+ # Output something different from input so the script gets removed
-+ cat "$1" >/dev/null
-+ EOF
-+
+ # Enable retries
+ test_config http.maxRetries 3 &&
+
+ # Git should retry after receiving 429 and eventually succeed
-+ git ls-remote "$HTTPD_URL/one_time_script/repo.git" >output 2>err &&
-+ test_grep "refs/heads/" output &&
-+
-+ # The one-time script should have been consumed (proving retry happened)
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ git ls-remote "$HTTPD_URL/http_429/retry-succeeds/1/repo.git" >output 2>err &&
++ test_grep "refs/heads/" output
+'
+
+test_expect_success 'HTTP 429 without Retry-After uses configured default' '
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF &&
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Rate limited - no retry info\n"
-+ cat "$1" >/dev/null
-+ EOF
-+
+ # Enable retries and configure default delay
+ test_config http.maxRetries 3 &&
+ test_config http.retryAfter 1 &&
+
+ # Git should retry using configured default and succeed
-+ git ls-remote "$HTTPD_URL/one_time_script/repo.git" >output 2>err &&
-+ test_grep "refs/heads/" output &&
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ git ls-remote "$HTTPD_URL/http_429/no-retry-after-header/none/repo.git" >output 2>err &&
++ test_grep "refs/heads/" output
+'
+
+test_expect_success 'HTTP 429 retry delays are respected' '
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF &&
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Retry-After: 2\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Rate limited\n"
-+ cat "$1" >/dev/null
-+ EOF
-+
+ # Enable retries
+ test_config http.maxRetries 3 &&
+
+ # Time the operation - it should take at least 2 seconds due to retry delay
+ start=$(date +%s) &&
-+ git ls-remote "$HTTPD_URL/one_time_script/repo.git" >output 2>err &&
++ git ls-remote "$HTTPD_URL/http_429/retry-delays-respected/2/repo.git" >output 2>err &&
+ end=$(date +%s) &&
+ duration=$((end - start)) &&
+
+ # Verify it took at least 2 seconds (allowing some tolerance)
+ test "$duration" -ge 1 &&
-+ test_grep "refs/heads/" output &&
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ test_grep "refs/heads/" output
+'
+
+test_expect_success 'HTTP 429 fails immediately if Retry-After exceeds http.maxRetryTime' '
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF &&
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Retry-After: 100\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Rate limited with long delay\n"
-+ cat "$1" >/dev/null
-+ EOF
-+
+ # Configure max retry time to 3 seconds (much less than requested 100)
+ test_config http.maxRetries 3 &&
+ test_config http.maxRetryTime 3 &&
+
+ # Should fail immediately without waiting
+ start=$(date +%s) &&
-+ test_must_fail git ls-remote "$HTTPD_URL/one_time_script/repo.git" 2>err &&
++ test_must_fail git ls-remote "$HTTPD_URL/http_429/retry-after-exceeds-max-time/100/repo.git" 2>err &&
+ end=$(date +%s) &&
+ duration=$((end - start)) &&
+
+ # Should fail quickly (less than 2 seconds, no 100 second wait)
+ test "$duration" -lt 2 &&
-+ test_grep "exceeds http.maxRetryTime" err &&
-+
-+ # The one-time script will be consumed on first request
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ test_grep "exceeds http.maxRetryTime" err
+'
+
+test_expect_success 'HTTP 429 fails if configured http.retryAfter exceeds http.maxRetryTime' '
+ # Test misconfiguration: retryAfter > maxRetryTime
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF &&
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Rate limited without header\n"
-+ cat "$1" >/dev/null
-+ EOF
-+
+ # Configure retryAfter larger than maxRetryTime
+ test_config http.maxRetries 3 &&
+ test_config http.retryAfter 100 &&
@@ t/t5584-http-429-retry.sh (new)
+
+ # Should fail immediately with configuration error
+ start=$(date +%s) &&
-+ test_must_fail git ls-remote "$HTTPD_URL/one_time_script/repo.git" 2>err &&
++ test_must_fail git ls-remote "$HTTPD_URL/http_429/config-retry-after-exceeds-max-time/none/repo.git" 2>err &&
+ end=$(date +%s) &&
+ duration=$((end - start)) &&
+
@@ t/t5584-http-429-retry.sh (new)
+ test_done
+ fi &&
+
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-EOF &&
-+ printf "Status: 429 Too Many Requests\\r\\n"
-+ printf "Retry-After: $future_date\\r\\n"
-+ printf "Content-Type: text/plain\\r\\n"
-+ printf "\\r\\n"
-+ printf "Rate limited with HTTP-date\\n"
-+ cat "\$1" >/dev/null
-+ EOF
++ # URL-encode the date (replace spaces with %20)
++ future_date_encoded=$(echo "$future_date" | sed "s/ /%20/g") &&
+
+ # Enable retries
+ test_config http.maxRetries 3 &&
+
+ # Git should parse the HTTP-date and retry after the delay
+ start=$(date +%s) &&
-+ git ls-remote "$HTTPD_URL/one_time_script/repo.git" >output 2>err &&
++ git ls-remote "$HTTPD_URL/http_429/http-date-format/$future_date_encoded/repo.git" >output 2>err &&
+ end=$(date +%s) &&
+ duration=$((end - start)) &&
+
+ # Should take at least 1 second (allowing tolerance for processing time)
+ test "$duration" -ge 1 &&
-+ test_grep "refs/heads/" output &&
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ test_grep "refs/heads/" output
+'
+
+test_expect_success 'HTTP 429 with HTTP-date exceeding maxRetryTime fails immediately' '
@@ t/t5584-http-429-retry.sh (new)
+ test_done
+ fi &&
+
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-EOF &&
-+ printf "Status: 429 Too Many Requests\\r\\n"
-+ printf "Retry-After: $future_date\\r\\n"
-+ printf "Content-Type: text/plain\\r\\n"
-+ printf "\\r\\n"
-+ printf "Rate limited with long HTTP-date\\n"
-+ cat "\$1" >/dev/null
-+ EOF
++ # URL-encode the date (replace spaces with %20)
++ future_date_encoded=$(echo "$future_date" | sed "s/ /%20/g") &&
+
+ # Configure max retry time much less than the 200 second delay
+ test_config http.maxRetries 3 &&
@@ t/t5584-http-429-retry.sh (new)
+
+ # Should fail immediately without waiting 200 seconds
+ start=$(date +%s) &&
-+ test_must_fail git ls-remote "$HTTPD_URL/one_time_script/repo.git" 2>err &&
++ test_must_fail git ls-remote "$HTTPD_URL/http_429/http-date-exceeds-max-time/$future_date_encoded/repo.git" 2>err &&
+ end=$(date +%s) &&
+ duration=$((end - start)) &&
+
+ # Should fail quickly (not wait 200 seconds)
+ test "$duration" -lt 2 &&
-+ test_grep "exceeds http.maxRetryTime" err &&
-+
-+ # The one-time script will be consumed on first request
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ test_grep "exceeds http.maxRetryTime" err
+'
+
+test_expect_success 'HTTP 429 with past HTTP-date should not wait' '
@@ t/t5584-http-429-retry.sh (new)
+ test_done
+ fi &&
+
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-EOF &&
-+ printf "Status: 429 Too Many Requests\\r\\n"
-+ printf "Retry-After: $past_date\\r\\n"
-+ printf "Content-Type: text/plain\\r\\n"
-+ printf "\\r\\n"
-+ printf "Rate limited with past date\\n"
-+ cat "\$1" >/dev/null
-+ EOF
++ # URL-encode the date (replace spaces with %20)
++ past_date_encoded=$(echo "$past_date" | sed "s/ /%20/g") &&
+
+ # Enable retries
+ test_config http.maxRetries 3 &&
+
+ # Git should retry immediately without waiting
+ start=$(date +%s) &&
-+ git ls-remote "$HTTPD_URL/one_time_script/repo.git" >output 2>err &&
++ git ls-remote "$HTTPD_URL/http_429/past-http-date/$past_date_encoded/repo.git" >output 2>err &&
+ end=$(date +%s) &&
+ duration=$((end - start)) &&
+
+ # Should complete quickly (less than 2 seconds)
+ test "$duration" -lt 2 &&
-+ test_grep "refs/heads/" output &&
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ test_grep "refs/heads/" output
+'
+
+test_expect_success 'HTTP 429 with invalid Retry-After format uses configured default' '
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF &&
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Retry-After: invalid-format-123abc\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Rate limited with malformed header\n"
-+ cat "$1" >/dev/null
-+ EOF
-+
+ # Configure default retry-after
+ test_config http.maxRetries 3 &&
+ test_config http.retryAfter 1 &&
+
+ # Should use configured default (1 second) since header is invalid
+ start=$(date +%s) &&
-+ git ls-remote "$HTTPD_URL/one_time_script/repo.git" >output 2>err &&
++ git ls-remote "$HTTPD_URL/http_429/invalid-retry-after-format/invalid/repo.git" >output 2>err &&
+ end=$(date +%s) &&
+ duration=$((end - start)) &&
+
+ # Should take at least 1 second (the configured default)
+ test "$duration" -ge 1 &&
+ test_grep "refs/heads/" output &&
-+ test_grep "waiting.*retry" err &&
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ test_grep "waiting.*retry" err
+'
+
+test_expect_success 'HTTP 429 will not be retried without config' '
+ # Default config means http.maxRetries=0 (retries disabled)
+ # When 429 is received, it should fail immediately without retry
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF &&
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Retry-After: 1\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Rate limited\n"
-+ cat "$1" >/dev/null
-+ EOF
-+
+ # Do NOT configure anything - use defaults (http.maxRetries defaults to 0)
+
+ # Should fail immediately without retry
-+ test_must_fail git ls-remote "$HTTPD_URL/one_time_script/repo.git" 2>err &&
++ test_must_fail git ls-remote "$HTTPD_URL/http_429/no-retry-without-config/1/repo.git" 2>err &&
+
+ # Verify no retry happened (no "waiting" message)
-+ ! grep -i "waiting.*retry" err &&
++ test_grep ! -i "waiting.*retry" err &&
+
+ # Should get 429 error
-+ test_grep "429" err &&
-+
-+ # The one-time script should be consumed on first request
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ test_grep "429" err
+'
+
+test_expect_success 'GIT_HTTP_RETRY_AFTER overrides http.retryAfter config' '
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF &&
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Rate limited - no Retry-After header\n"
-+ cat "$1" >/dev/null
-+ EOF
-+
+ # Configure retryAfter to 10 seconds
+ test_config http.maxRetries 3 &&
+ test_config http.retryAfter 10 &&
+
+ # Override with environment variable to 1 second
+ start=$(date +%s) &&
-+ GIT_HTTP_RETRY_AFTER=1 git ls-remote "$HTTPD_URL/one_time_script/repo.git" >output 2>err &&
++ GIT_HTTP_RETRY_AFTER=1 git ls-remote "$HTTPD_URL/http_429/env-retry-after-override/none/repo.git" >output 2>err &&
+ end=$(date +%s) &&
+ duration=$((end - start)) &&
+
@@ t/t5584-http-429-retry.sh (new)
+ test "$duration" -ge 1 &&
+ test "$duration" -lt 5 &&
+ test_grep "refs/heads/" output &&
-+ test_grep "waiting.*retry" err &&
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ test_grep "waiting.*retry" err
+'
+
+test_expect_success 'GIT_HTTP_MAX_RETRIES overrides http.maxRetries config' '
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF &&
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Retry-After: 1\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Rate limited\n"
-+ cat "$1" >/dev/null
-+ EOF
-+
+ # Configure maxRetries to 0 (disabled)
+ test_config http.maxRetries 0 &&
+ test_config http.retryAfter 1 &&
+
+ # Override with environment variable to enable retries
-+ GIT_HTTP_MAX_RETRIES=3 git ls-remote "$HTTPD_URL/one_time_script/repo.git" >output 2>err &&
++ GIT_HTTP_MAX_RETRIES=3 git ls-remote "$HTTPD_URL/http_429/env-max-retries-override/1/repo.git" >output 2>err &&
+
+ # Should retry (env var enables it despite config saying disabled)
+ test_grep "refs/heads/" output &&
-+ test_grep "waiting.*retry" err &&
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ test_grep "waiting.*retry" err
+'
+
+test_expect_success 'GIT_HTTP_MAX_RETRY_TIME overrides http.maxRetryTime config' '
-+ write_script "$HTTPD_ROOT_PATH/one-time-script" <<-\EOF &&
-+ printf "Status: 429 Too Many Requests\r\n"
-+ printf "Retry-After: 50\r\n"
-+ printf "Content-Type: text/plain\r\n"
-+ printf "\r\n"
-+ printf "Rate limited with long delay\n"
-+ cat "$1" >/dev/null
-+ EOF
-+
+ # Configure maxRetryTime to 100 seconds (would accept 50 second delay)
+ test_config http.maxRetries 3 &&
+ test_config http.maxRetryTime 100 &&
@@ t/t5584-http-429-retry.sh (new)
+ # Override with environment variable to 10 seconds (should reject 50 second delay)
+ start=$(date +%s) &&
+ test_must_fail env GIT_HTTP_MAX_RETRY_TIME=10 \
-+ git ls-remote "$HTTPD_URL/one_time_script/repo.git" 2>err &&
++ git ls-remote "$HTTPD_URL/http_429/env-max-retry-time-override/50/repo.git" 2>err &&
+ end=$(date +%s) &&
+ duration=$((end - start)) &&
+
+ # Should fail quickly (not wait 50 seconds) because env var limits to 10
+ test "$duration" -lt 5 &&
-+ test_grep "exceeds http.maxRetryTime" err &&
-+ test_path_is_missing "$HTTPD_ROOT_PATH/one-time-script"
++ test_grep "exceeds http.maxRetryTime" err
+'
+
+test_expect_success 'verify normal repository access still works' '
2: 4382237922 < -: ---------- remote-curl: fix memory leak in show_http_message()
3: adbcc0251f ! 2: ad4495fc94 http: add trace2 logging for retry operations @@ http.c
static struct trace_key trace_curl = TRACE_KEY_INIT(CURL);
static int trace_curl_data = 1;
+@@ http.c: void run_active_slot(struct active_request_slot *slot)
+
+ if (waiting_for_delay) {
+ warning(_("rate limited, waiting %ld seconds before retry"), slot->retry_delay_seconds);
++ trace2_data_intmax("http", the_repository, "http/retry-sleep-seconds",
++ slot->retry_delay_seconds);
+ start_time = slot->retry_delay_start;
+ }
+
+@@ http.c: void run_active_slot(struct active_request_slot *slot)
+ }
+
+ if (elapsed_time.tv_sec >= slot->retry_delay_seconds) {
++ trace2_region_leave("http", "retry-sleep", the_repository);
+ slot->retry_delay_seconds = -1;
+ waiting_for_delay = 0;
+
@@ http.c: static int handle_curl_result(struct slot_results *results)
+ return HTTP_REAUTH;
+ }
} else if (results->http_code == 429) {
- /* Store the retry_after value for use in retry logic */
- last_retry_after = results->retry_after;
+ trace2_data_intmax("http", the_repository, "http/429-retry-after",
-+ last_retry_after);
++ results->retry_after);
return HTTP_RATE_LIMITED;
} else {
if (results->http_connectcode == 407)
-@@ http.c: static void sleep_for_retry(long retry_after)
- if (retry_after > 0) {
- unsigned int remaining;
- warning(_("rate limited, waiting %ld seconds before retry"), retry_after);
-+ trace2_region_enter("http", "retry-sleep", the_repository);
-+ trace2_data_intmax("http", the_repository, "http/retry-sleep-seconds",
-+ retry_after);
- remaining = sleep(retry_after);
- while (remaining > 0) {
- /* Sleep was interrupted, continue sleeping */
- remaining = sleep(remaining);
+@@ http.c: static void sleep_for_retry(struct active_request_slot *slot, long retry_after)
+ static long handle_rate_limit_retry(int *rate_limit_retries, long slot_retry_after)
+ {
+ int retry_attempt = http_max_retries - *rate_limit_retries + 1;
++
++ trace2_data_intmax("http", the_repository, "http/429-retry-attempt",
++ retry_attempt);
++
+ if (*rate_limit_retries <= 0) {
+ /* Retries are disabled or exhausted */
+ if (http_max_retries > 0) {
+ error(_("too many rate limit retries, giving up"));
++ trace2_data_string("http", the_repository,
++ "http/429-error", "retries-exhausted");
}
-+ trace2_region_leave("http", "retry-sleep", the_repository);
+ return -1;
}
- }
-
-@@ http.c: static int http_request_reauth(const char *url,
- /* Handle rate limiting with retry logic */
- int retry_attempt = http_max_retries - rate_limit_retries + 1;
-
-+ trace2_data_intmax("http", the_repository, "http/429-retry-attempt",
-+ retry_attempt);
-+
- if (rate_limit_retries <= 0) {
- /* Retries are disabled or exhausted */
- if (http_max_retries > 0) {
- error(_("too many rate limit retries, giving up"));
-+ trace2_data_string("http", the_repository,
-+ "http/429-error", "retries-exhausted");
- }
- return HTTP_ERROR;
- }
-@@ http.c: static int http_request_reauth(const char *url,
- error(_("rate limited (HTTP 429) requested %ld second delay, "
- "exceeds http.maxRetryTime of %ld seconds"),
- last_retry_after, http_max_retry_time);
-+ trace2_data_string("http", the_repository,
-+ "http/429-error", "exceeds-max-retry-time");
-+ trace2_data_intmax("http", the_repository,
-+ "http/429-requested-delay", last_retry_after);
- last_retry_after = -1; /* Reset after use */
- return HTTP_ERROR;
- }
-@@ http.c: static int http_request_reauth(const char *url,
- /* Not configured - exit with error */
- error(_("rate limited (HTTP 429) and no Retry-After header provided. "
- "Configure http.retryAfter or set GIT_HTTP_RETRY_AFTER."));
-+ trace2_data_string("http", the_repository,
-+ "http/429-error", "no-retry-after-config");
- return HTTP_ERROR;
- }
-- /* Check if configured default exceeds maximum allowed */
-- if (http_retry_after > http_max_retry_time) {
-- error(_("configured http.retryAfter (%ld seconds) exceeds "
-- "http.maxRetryTime (%ld seconds)"),
-- http_retry_after, http_max_retry_time);
-- return HTTP_ERROR;
-- }
-- /* Use configured default retry-after value */
-- sleep_for_retry(http_retry_after);
-+ /* Check if configured default exceeds maximum allowed */
-+ if (http_retry_after > http_max_retry_time) {
-+ error(_("configured http.retryAfter (%ld seconds) exceeds "
-+ "http.maxRetryTime (%ld seconds)"),
-+ http_retry_after, http_max_retry_time);
-+ trace2_data_string("http", the_repository,
-+ "http/429-error", "config-exceeds-max-retry-time");
-+ return HTTP_ERROR;
-+ }
-+ /* Use configured default retry-after value */
+@@ http.c: static long handle_rate_limit_retry(int *rate_limit_retries, long slot_retry_aft
+ error(_("rate limited (HTTP 429) requested %ld second delay, "
+ "exceeds http.maxRetryTime of %ld seconds"),
+ slot_retry_after, http_max_retry_time);
+ trace2_data_string("http", the_repository,
-+ "http/429-retry-source", "config-default");
-+ sleep_for_retry(http_retry_after);
- }
- } else if (ret == HTTP_REAUTH) {
- credential_fill(the_repository, &http_auth, 1);
++ "http/429-error", "exceeds-max-retry-time");
++ trace2_data_intmax("http", the_repository,
++ "http/429-requested-delay", slot_retry_after);
+ return -1;
+ }
+ return slot_retry_after;
+@@ http.c: static long handle_rate_limit_retry(int *rate_limit_retries, long slot_retry_aft
+ /* Not configured - exit with error */
+ error(_("rate limited (HTTP 429) and no Retry-After header provided. "
+ "Configure http.retryAfter or set GIT_HTTP_RETRY_AFTER."));
++ trace2_data_string("http", the_repository,
++ "http/429-error", "no-retry-after-config");
+ return -1;
+ }
+ /* Check if configured default exceeds maximum allowed */
+@@ http.c: static long handle_rate_limit_retry(int *rate_limit_retries, long slot_retry_aft
+ error(_("configured http.retryAfter (%ld seconds) exceeds "
+ "http.maxRetryTime (%ld seconds)"),
+ http_retry_after, http_max_retry_time);
++ trace2_data_string("http", the_repository,
++ "http/429-error", "config-exceeds-max-retry-time");
+ return -1;
+ }
+-
++ trace2_data_string("http", the_repository,
++ "http/429-retry-source", "config-default");
+ return http_retry_after;
+ }
+ }
Signature · 1 line
gitgitgadget
PATCH v2 1/2 http: add support for HTTP 429 rate limit retries
From: Vaidas Pilkauskas <[email protected]>
Add retry logic for HTTP 429 (Too Many Requests) responses to handle server-side rate limiting gracefully. When Git's HTTP client receives a 429 response, it can now automatically retry the request after an appropriate delay, respecting the server's rate limits.
The implementation supports the RFC-compliant Retry-After header in both delay-seconds (integer) and HTTP-date (RFC 2822) formats. If a past date is provided, Git retries immediately without waiting.
Retry behavior is controlled by three new configuration options (http.maxRetries, http.retryAfter, and http.maxRetryTime) which are documented in git-config(1).
The retry logic implements a fail-fast approach: if any delay (whether from server header or configuration) exceeds maxRetryTime, Git fails immediately with a clear error message rather than capping the delay. This provides better visibility into rate limiting issues.
The implementation includes extensive test coverage for basic retry behavior, Retry-After header formats (integer and HTTP-date), configuration combinations, maxRetryTime limits, invalid header handling, environment variable overrides, and edge cases.
--- Documentation/config/http.adoc | 24 +++ http-push.c | 8 + http-walker.c | 5 + http.c | 300 ++++++++++++++++++++++++++++----- http.h | 4 + remote-curl.c | 49 ++++-- strbuf.c | 2 +- t/lib-httpd.sh | 1 + t/lib-httpd/apache.conf | 8 + t/lib-httpd/http-429.sh | 98 +++++++++++ t/meson.build | 1 + t/t5584-http-429-retry.sh | 286 +++++++++++++++++++++++++++++++ 12 files changed, 729 insertions(+), 57 deletions(-) create mode 100644 t/lib-httpd/http-429.sh create mode 100755 t/t5584-http-429-retry.sh
Documentation/config/http.adoc +24 -0
@@ -315,6 +315,30 @@ http.keepAliveCount:: unset, curl's default value is used. Can be overridden by the `GIT_HTTP_KEEPALIVE_COUNT` environment variable. +http.retryAfter:: + Default wait time in seconds before retrying when a server returns + HTTP 429 (Too Many Requests) without a Retry-After header. If set + to -1 (the default), Git will fail immediately when encountering + a 429 response without a Retry-After header. When a Retry-After + header is present, its value takes precedence over this setting. + Can be overridden by the `GIT_HTTP_RETRY_AFTER` environment variable. + See also `http.maxRetries` and `http.maxRetryTime`. + +http.maxRetries:: + Maximum number of times to retry after receiving HTTP 429 (Too Many + Requests) responses. Set to 0 (the default) to disable retries. + Can be overridden by the `GIT_HTTP_MAX_RETRIES` environment variable. + See also `http.retryAfter` and `http.maxRetryTime`. + +http.maxRetryTime:: + Maximum time in seconds to wait for a single retry attempt when + handling HTTP 429 (Too Many Requests) responses. If the server + requests a delay (via Retry-After header) or if `http.retryAfter` + is configured with a value that exceeds this maximum, Git will fail + immediately rather than waiting. Default is 300 seconds (5 minutes). + Can be overridden by the `GIT_HTTP_MAX_RETRY_TIME` environment + variable. See also `http.retryAfter` and `http.maxRetries`. + http.noEPSV:: A boolean which disables using of EPSV ftp command by curl. This can be helpful with some "poor" ftp servers which don't
http-push.c +8 -0
@@ -716,6 +716,10 @@ static int fetch_indices(void)case HTTP_MISSING_TARGET : ret =0 ;break ; +case HTTP_RATE_LIMITED : +error (_ ("rate limited by '%s', please try again later" ), repo->url); + ret =-1 ; +break ;default : ret =-1 ; }
@@ -1548,6 +1552,10 @@ static int remote_exists(const char *path)case HTTP_MISSING_TARGET : ret =0 ;break ; +case HTTP_RATE_LIMITED : +error (_ ("rate limited by '%s', please try again later" ), url); + ret =-1 ; +break ;case HTTP_ERROR :error ("unable to access '%s': %s" , url, curl_errorstr);/* fallthrough */
http-walker.c +5 -0
@@ -414,6 +414,11 @@ static int fetch_indices(struct walker *walker, struct alt_base *repo) repo->got_indices =1 ; ret =0 ;break ; +case HTTP_RATE_LIMITED : +error ("rate limited by '%s', please try again later" , repo->base); + repo->got_indices =0 ; + ret =-1 ; +break ;default : repo->got_indices =0 ; ret =-1 ;
http.c +261 -39
446 lines not on this page · raw message
http.h +4 -0
@@ -20,6 +20,7 @@ struct slot_results {long http_code;long auth_avail;long http_connectcode; +long retry_after; };struct active_request_slot {
@@ -32,6 +33,8 @@ struct active_request_slot {void *callback_data;void (*callback_func)(void *data);struct active_request_slot *next; +long retry_delay_seconds; +struct timeval retry_delay_start; };struct buffer {
@@ -167,6 +170,7 @@ struct http_get_options {#define HTTP_REAUTH 4#define HTTP_NOAUTH 5#define HTTP_NOMATCHPUBLICKEY 6 +#define HTTP_RATE_LIMITED 7/* * Requests a URL and stores the result in a strbuf.
remote-curl.c +32 -17
83 lines not on this page · raw message
strbuf.c +1 -1
@@ -168,7 +168,7 @@ int strbuf_reencode(struct strbuf *sb, const char *from, const char *to)if (!out)return -1 ; -strbuf_attach (sb, out, len, len); +strbuf_attach (sb, out, len, len +1 );return 0 ; }
t/lib-httpd.sh +1 -0
@@ -167,6 +167,7 @@ prepare_httpd() { install_script error.sh install_script apply-one-time-script.sh install_script nph-custom-auth.sh + install_script http-429.sh ln -s "$LIB_HTTPD_MODULE_PATH" "$HTTPD_ROOT_PATH/modules"
t/lib-httpd/apache.conf +8 -0
@@ -139,6 +139,10 @@ SetEnv PERL_PATH ${PERL_PATH} SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} SetEnv GIT_HTTP_EXPORT_ALL </LocationMatch> +<LocationMatch /http_429/> + SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} + SetEnv GIT_HTTP_EXPORT_ALL +</LocationMatch> <LocationMatch /smart_v0/> SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} SetEnv GIT_HTTP_EXPORT_ALL
@@ -160,6 +164,7 @@ ScriptAlias /broken_smart/ broken-smart-http.sh/ ScriptAlias /error_smart/ error-smart-http.sh/ ScriptAlias /error/ error.sh/ ScriptAliasMatch /one_time_script/(.*) apply-one-time-script.sh/$1 +ScriptAliasMatch /http_429/(.*) http-429.sh/$1 ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1 <Directory ${GIT_EXEC_PATH}> Options FollowSymlinks
@@ -185,6 +190,9 @@ ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1 <Files apply-one-time-script.sh> Options ExecCGI </Files> +<Files http-429.sh> + Options ExecCGI +</Files> <Files ${GIT_EXEC_PATH}/git-http-backend> Options ExecCGI </Files>
t/lib-httpd/http-429.sh new file +98 -0
98 lines not on this page · raw message
t/meson.build +1 -0
@@ -700,6 +700,7 @@ integration_tests = [ 't5581-http-curl-verbose.sh', 't5582-fetch-negative-refspec.sh', 't5583-push-branches.sh', + 't5584-http-429-retry.sh', 't5600-clone-fail-cleanup.sh', 't5601-clone.sh', 't5602-clone-remote-exec.sh',
t/t5584-http-429-retry.sh new file +286 -0
286 lines not on this page · raw message
Signature · 1 line
gitgitgadget
PATCH v2 2/2 http: add trace2 logging for retry operations
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
Add trace2 instrumentation to HTTP 429 retry operations to enable monitoring and debugging of rate limit scenarios in production environments.
The trace2 logging captures:
* Retry attempt numbers (http/429-retry-attempt) to track retry progression and identify how many attempts were needed
* Retry-After header values (http/429-retry-after) from server responses to understand server-requested delays
* Actual sleep durations (http/retry-sleep-seconds) within trace2 regions (http/retry-sleep) to measure time spent waiting
* Error conditions (http/429-error) such as "retries-exhausted", "exceeds-max-retry-time", "no-retry-after-config", and "config-exceeds-max-retry-time" for diagnosing failures
* Retry source (http/429-retry-source) indicating whether delay came from server header or config default
This instrumentation provides complete visibility into retry behavior, enabling operators to monitor rate limiting patterns, diagnose retry failures, and optimize retry configuration based on real-world data.
--- http.c | 23 ++++++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-)
http.c +22 -1
@@ -23,6 +23,7 @@#include "odb.h" #include "tempfile.h" #include "date.h" +#include "trace2.h" static struct trace_key trace_curl =TRACE_KEY_INIT (CURL );static int trace_curl_data =1 ;
@@ -1738,6 +1739,8 @@ void run_active_slot(struct active_request_slot *slot)if (waiting_for_delay) {warning (_ ("rate limited, waiting %ld seconds before retry" ), slot->retry_delay_seconds); +trace2_data_intmax ("http" , the_repository,"http/retry-sleep-seconds" , + slot->retry_delay_seconds); start_time = slot->retry_delay_start; }
@@ -1753,6 +1756,7 @@ void run_active_slot(struct active_request_slot *slot) }if (elapsed_time.tv_sec >= slot->retry_delay_seconds) { +trace2_region_leave ("http" ,"retry-sleep" , the_repository); slot->retry_delay_seconds =-1 ; waiting_for_delay =0 ;
@@ -1995,6 +1999,8 @@ static int handle_curl_result(struct slot_results *results)return HTTP_REAUTH ; } }else if (results->http_code ==429 ) { +trace2_data_intmax ("http" , the_repository,"http/429-retry-after" , + results->retry_after);return HTTP_RATE_LIMITED ; }else {if (results->http_connectcode ==407 )
@@ -2421,10 +2427,16 @@ static void sleep_for_retry(struct active_request_slot *slot, long retry_after)static long handle_rate_limit_retry (int *rate_limit_retries,long slot_retry_after) {int retry_attempt = http_max_retries - *rate_limit_retries +1 ; + +trace2_data_intmax ("http" , the_repository,"http/429-retry-attempt" , + retry_attempt); +if (*rate_limit_retries <=0 ) {/* Retries are disabled or exhausted */ if (http_max_retries >0 ) {error (_ ("too many rate limit retries, giving up" )); +trace2_data_string ("http" , the_repository, +"http/429-error" ,"retries-exhausted" ); }return -1 ; }
@@ -2439,6 +2451,10 @@ static long handle_rate_limit_retry(int *rate_limit_retries, long slot_retry_afterror (_ ("rate limited (HTTP 429) requested %ld second delay, " "exceeds http.maxRetryTime of %ld seconds" ), slot_retry_after, http_max_retry_time); +trace2_data_string ("http" , the_repository, +"http/429-error" ,"exceeds-max-retry-time" ); +trace2_data_intmax ("http" , the_repository, +"http/429-requested-delay" , slot_retry_after);return -1 ; }return slot_retry_after;
@@ -2448,6 +2464,8 @@ static long handle_rate_limit_retry(int *rate_limit_retries, long slot_retry_aft/* Not configured - exit with error */ error (_ ("rate limited (HTTP 429) and no Retry-After header provided. " "Configure http.retryAfter or set GIT_HTTP_RETRY_AFTER." )); +trace2_data_string ("http" , the_repository, +"http/429-error" ,"no-retry-after-config" );return -1 ; }/* Check if configured default exceeds maximum allowed */
@@ -2455,9 +2473,12 @@ static long handle_rate_limit_retry(int *rate_limit_retries, long slot_retry_afterror (_ ("configured http.retryAfter (%ld seconds) exceeds " "http.maxRetryTime (%ld seconds)" ), http_retry_after, http_max_retry_time); +trace2_data_string ("http" , the_repository, +"http/429-error" ,"config-exceeds-max-retry-time" );return -1 ; } - +trace2_data_string ("http" , the_repository, +"http/429-retry-source" ,"config-default" );return http_retry_after; } }
Signature · 1 line
gitgitgadget
Quoted 28 lines · Vaidas Pilkauskas via GitGitGadget
From: Vaidas Pilkauskas <[email protected]>
Add retry logic for HTTP 429 (Too Many Requests) responses to handle server-side rate limiting gracefully. When Git's HTTP client receives a 429 response, it can now automatically retry the request after an appropriate delay, respecting the server's rate limits.
The implementation supports the RFC-compliant Retry-After header in both delay-seconds (integer) and HTTP-date (RFC 2822) formats. If a past date is provided, Git retries immediately without waiting.
Retry behavior is controlled by three new configuration options (http.maxRetries, http.retryAfter, and http.maxRetryTime) which are documented in git-config(1).
The retry logic implements a fail-fast approach: if any delay (whether from server header or configuration) exceeds maxRetryTime, Git fails immediately with a clear error message rather than capping the delay. This provides better visibility into rate limiting issues.
The implementation includes extensive test coverage for basic retry behavior, Retry-After header formats (integer and HTTP-date), configuration combinations, maxRetryTime limits, invalid header handling, environment variable overrides, and edge cases.
Documentation/config/http.adoc http.keepAliveCount:: · line 321
Quoted 4 lines
Documentation/config/http.adoc +4
-0+http.retryAfter:: + Default wait time in seconds before retrying when a server returns + HTTP 429 (Too Many Requests) without a Retry-After header. If set + to -1 (the default), Git will fail immediately when encountering
While reviewing, I originally wrote:
Setting the default as "-1" makes sense to me. The current behavior is to give up when we receive a HTTP 429 response with or without a Retry-After header, so retaining that behavior makes sense and seems like a sensible path.
, but I'm not sure that I am sold on that line of thinking. This is controlling how long we'll wait after a 429 response before retrying, not how many times we'll retry (which is http.maxRetries below).
Should the default here be zero? We would "retry" immediately, but that retry would fail since the maximum retries is set to "zero" by default.
Quoted 16 lines
http-push.c +4
-0@@ -716,6 +716,10 @@ static int fetch_indices(void)case HTTP_MISSING_TARGET : ret =0 ;break ; +case HTTP_RATE_LIMITED : +error (_ ("rate limited by '%s', please try again later" ), repo->url); + ret =-1 ; +break ;default : ret =-1 ; }
http-push.c static int remote_exists(const char *path) · line 1558
Quoted 7 lines
http-push.c +4
-0@@ -1548,6 +1552,10 @@ static int remote_exists(const char *path)case HTTP_MISSING_TARGET : ret =0 ;break ; +case HTTP_RATE_LIMITED : +error (_ ("rate limited by '%s', please try again later" ), url); + ret =-1 ; +break ;
I wonder if there is an opportunity to DRY this up a bit? I think the case in fetch_indices() is very similar to remote_Exists(), and ditto for fetch_indices() in the http-walker.c code.
The only exception I could see is http-walker.c's fetch_indices() needs to also set repo->got_indices, but I think that could be done as a separate pass.
If you end up going in that direction, I would suggest pulling out a function as a preparatory commit before introducing the changes in this patch so that you when you are ready to add the "rate limited by '%s'" error(), you only have to do so once.
http-push.c static int remote_exists(const char *path) · line 1559
Quoted 2 lines
http-push.c +1
-1-static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UNUSED) +static size_t fwrite_headers (char *ptr,size_t eltsize,size_t nmemb,void *p)
Thanks for making this change. I think that handling both www-authenticate and retry-after headers in the same function makes a lot of sense, and the new name reflects that appropriately.
Quoted 20 lines
http-push.c +0
-0{size_t size = eltsize * nmemb;http-push.c +7
-0struct strvec *values = &http_auth.wwwauth_headers;struct strbuf buf =STRBUF_INIT ;const char *val;size_t val_len; +struct active_request_slot *slot = (struct active_request_slot *)p;/* * Header lines may not come NULL-terminated from libcurl so we must @@ -257,6 +264,47 @@ static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UN goto exit; } +/* Parse Retry-After header for rate limiting */ +if (skip_iprefix_mem (ptr, size,"retry-after:" , &val, &val_len)) { +strbuf_add (&buf, val, val_len); +strbuf_trim (&buf); + +if (slot && slot->results) {
http-push.c static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UN · line 282
Quoted 10 lines
http-push.c +10
-0+/* Parse the retry-after value (delay-seconds or HTTP-date) */ +char *endptr; +long retry_after; + + errno =0 ; + retry_after =strtol (buf.buf, &endptr,10 ); + +/* Check if it's a valid integer (delay-seconds format) */ +if (endptr != buf.buf && *endptr =='\0' && + errno !=ERANGE && retry_after >0 ) {
Should we handle "Retry-After: 0" here? I think that this means "retry immediately", so I imagine that we should change this to read "&& retry_after >= 0" instead.
Quoted 12 lines
http-push.c +12
-0+ slot->results->retry_after = retry_after; + }else { +/* Try parsing as HTTP-date format */ +timestamp_t timestamp; +int offset; +if (!parse_date_basic (buf.buf, ×tamp, &offset)) { +/* Successfully parsed as date, calculate delay from now */ +timestamp_t now =time (NULL ); +if (timestamp > now) { + slot->results->retry_after = (long )(timestamp - now); + }else { +/* Past date means retry immediately */
http-push.c static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UN · line 304
Quoted 10 lines
http-push.c +10
-0+ slot->results->retry_after =0 ; + } + }else { +/* Failed to parse as either delay-seconds or HTTP-date */ +warning (_ ("unable to parse Retry-After header value: '%s'" ), buf.buf); + } + } + } + + http_auth.header_is_last_match =1 ;
Could you help me understand why we're setting header_is_last_match here? I think since we immediately "goto exit" this line isn't strictly necessary.
As a separate but related note, I don't know if this function properly handles header continuations for Retry-After headers, but in practice I suspect it doesn't matter, as servers should not be continuing Retry-After headers across multiple lines.
Quoted 24 lines
http-push.c +16
-1@@ -1660,44 +1729,98 @@ void run_active_slot(struct active_request_slot *slot)fd_set excfds;int max_fd;struct timeval select_timeout; +long curl_timeout; +struct timeval start_time = {0 }, current_time, elapsed_time = {0 }; +long remaining_seconds;int finished =0 ; +int slot_not_started = (slot->finished ==NULL ); +int waiting_for_delay = (slot->retry_delay_seconds >0 ); + +if (waiting_for_delay) { +warning (_ ("rate limited, waiting %ld seconds before retry" ), slot->retry_delay_seconds); + start_time = slot->retry_delay_start; + } slot->finished = &finished; -while (!finished) { +while (waiting_for_delay || !finished) { +if (waiting_for_delay) { +gettimeofday (¤t_time,NULL ); + elapsed_time.tv_sec = current_time.tv_sec - start_time.tv_sec; + elapsed_time.tv_usec = current_time.tv_usec - start_time.tv_usec; +if (elapsed_time.tv_usec <0 ) {
http-push.c void run_active_slot(struct active_request_slot *slot) · line 1760
Quoted 10 lines
http-push.c +10
-0+ elapsed_time.tv_sec--; + elapsed_time.tv_usec +=1000000 ; + } + +if (elapsed_time.tv_sec >= slot->retry_delay_seconds) { + slot->retry_delay_seconds =-1 ; + waiting_for_delay =0 ; + +if (slot_not_started) +return ;
I wonder if run_active_slot() is the right place for these changes or if it should be handled separately. I think it may be somewhat surprising for run_active_slot() to return without actually running the slot, even if the slot is marked as "active" but just waiting for a delay.
OTOH, like I mentioned earlier, I am far from an expert in this part of the code, so perhaps this is totally OK. shortlog says that Peff (CC'd) is among the most active contributors to this file in the past year, so I'll be curious what he thinks as well.
Quoted 10 lines
http-push.c +2
-0@@ -1871,6 +1994,8 @@ static int handle_curl_result(struct slot_results *results) }return HTTP_REAUTH ; } + }else if (results->http_code ==429 ) { +return HTTP_RATE_LIMITED ; }else {if (results->http_connectcode ==407 )credential_reject (the_repository, &proxy_auth);
http-push.c int run_one_slot(struct active_request_slot *slot, · line 2020
Quoted 10 lines
http-push.c +7
-0@@ -1886,6 +2011,14 @@ int run_one_slot(struct active_request_slot *slot, struct slot_results *results) { slot->results = results; +/* Initialize retry_after to -1 (not set) */ + results->retry_after =-1 ; + +/* If there's a retry delay, wait for it before starting the slot */ +if (slot->retry_delay_seconds >0 ) { +run_active_slot (slot); + }
This is a nitpick, but the curly braces here are unnecessary for a single-line if statement. Documentation/CodingGuidelines has more details here.
Quoted 20 lines
http-push.c +6
-1+if (!start_active_slot (slot)) {xsnprintf (curl_errorstr,sizeof (curl_errorstr),"failed to start HTTP request" );@@ -2117,9 +2250,13 @@ static void http_opt_request_remainder(CURL *curl, off_t pos)#define HTTP_REQUEST_STRBUF 0#define HTTP_REQUEST_FILE 1 +static void sleep_for_retry (struct active_request_slot *slot,long retry_after); +static int http_request (const char *url,void *result,int target, -const struct http_get_options *options) +const struct http_get_options *options, +long *retry_after_out, +long retry_delay) {struct active_request_slot *slot;struct slot_results results;
http-push.c static int http_request(const char *url, · line 2272
Quoted 7 lines
http-push.c +4
-0@@ -2129,6 +2266,10 @@ static int http_request(const char *url,int ret; slot =get_active_slot (); +/* Mark slot for delay if retry delay is provided */ +if (retry_delay >0 ) { +sleep_for_retry (slot, retry_delay); + }
Same note here as above.
Quoted 4 lines
http-push.c +4
-0+/* +* Handle rate limiting retry logic for HTTP 429 responses. +* Uses slot-specific retry_after value to support concurrent slots. +* Returns a negative value if retries are exhausted or configuration is invalid,
http-push.c static int http_request(const char *url,
Quoted 10 lines
http-push.c +10
-0+* otherwise returns the delay value (>= 0) to indicate the retry should proceed. +*/ +static long handle_rate_limit_retry (int *rate_limit_retries,long slot_retry_after) +{ +int retry_attempt = http_max_retries - *rate_limit_retries +1 ; +if (*rate_limit_retries <=0 ) { +/* Retries are disabled or exhausted */ +if (http_max_retries >0 ) { +error (_ ("too many rate limit retries, giving up" )); + }
Here as well.
Quoted 21 lines
http-push.c +21
-0+return -1 ; + } + +/* Decrement retries counter */ + (*rate_limit_retries)--; + +/* Use the slot-specific retry_after value or configured default */ +if (slot_retry_after >=0 ) { +/* Check if retry delay exceeds maximum allowed */ +if (slot_retry_after > http_max_retry_time) { +error (_ ("rate limited (HTTP 429) requested %ld second delay, " +"exceeds http.maxRetryTime of %ld seconds" ), + slot_retry_after, http_max_retry_time); +return -1 ; + } +return slot_retry_after; + }else { +/* No Retry-After header provided */ +if (http_retry_after <0 ) { +/* Not configured - exit with error */ + error(_("rate limited (HTTP 429) and no Retry-After header provided. "
http-push.c
Quoted 10 lines
http-push.c +10
-0+ "Configure http.retryAfter or set GIT_HTTP_RETRY_AFTER.")); +return -1 ; + } +/* Check if configured default exceeds maximum allowed */ +if (http_retry_after > http_max_retry_time) { +error (_ ("configured http.retryAfter (%ld seconds) exceeds " +"http.maxRetryTime (%ld seconds)" ), + http_retry_after, http_max_retry_time); +return -1 ; + }
As a general note on these error()s, I wonder if it would be worth shortening them up a bit. For example, the first one reads:
"rate limited (HTTP 429) requested %ld second delay, exceeds http.maxRetryTime of %ld seconds"
Perhaps we could shorten this to something like:
"response requested a delay greater than http.maxRetryTime (%ld > %ld seconds)"
I feel like we could get it even shorter, but I think that this is a good starting point.
As an additional note, I think we generally try and avoid putting instructions like "Configure http.retryAfter or [...]" in error() messages. Those would be good advise() messages, enabling the user to turn them off if they are not relevant to their situation, whereas error() messages are fixed.
Quoted 11 lines
http-push.c +3
-0+static int http_request_recoverable (const char *url,void *result,int target,struct http_get_options *options) {int i =3 ;int ret; +int rate_limit_retries = http_max_retries; +long slot_retry_after =-1 ;/* Per-slot retry_after value */ if (always_auth_proactively ())credential_fill (the_repository, &http_auth,1 );
http-push.c
Quoted 10 lines
http-push.c +4
-2- ret =http_request (url, result, target, options); + ret =http_request (url, result, target, options, &slot_retry_after,-1 ); -if (ret !=HTTP_OK && ret !=HTTP_REAUTH ) +if (ret !=HTTP_OK && ret !=HTTP_REAUTH && ret !=HTTP_RATE_LIMITED )return ret; +/* If retries are disabled and we got a 429, fail immediately */ +if (ret ==HTTP_RATE_LIMITED && http_max_retries ==0 )
Another minor CodingGuidelines nit, but we generally do not write "x == 0", and instead prefer "!x".
http-push.c static int http_request_reauth(const char *url, · line 2494
Quoted 11 lines
http-push.c +3
-1+return HTTP_ERROR ; +if (options && options->effective_url && options->base_url) {if (update_url_from_redirect (options->base_url, url, options->effective_url)) {@@ -2276,7 +2491,8 @@ static int http_request_reauth(const char *url, } } -while (ret ==HTTP_REAUTH && --i) { +while ((ret ==HTTP_REAUTH || ret ==HTTP_RATE_LIMITED ) && --i) {
I had to re-read this line, since I wasn't sure that decrementing i was the right thing to do for both reauth and rate limited responses. But it is, since we pass a pointer to rate_limit_retries down to handle_rate_limit_retry() which will decrement it and eventually cause it to return -1 when retries are exhausted, causing this loop to exit.
http-push.c static struct discovery *discover_refs(const char *service, int for_push) · line 534
Quoted 11 lines
http-push.c +3
-3static int get_protocol_http_header(enum protocol_version version,@@ -518,21 +529,25 @@ static struct discovery *discover_refs(const char *service, int for_push)case HTTP_OK :break ;case HTTP_MISSING_TARGET : -show_http_message (&type, &charset, &buffer); -die (_ ("repository '%s' not found" ), -transport_anonymize_url (url.buf)); +show_http_message_fatal (&type, &charset, &buffer, +_ ("repository '%s' not found" ), +transport_anonymize_url (url.buf));
Thanks for taking my suggestion here as well. I think that the end result reads much cleaner, though I do think that introducing the new show_http_message_fatal() function and rewriting the existing code should happen in a preparatory commit before this one to more clearly separate the changes.
strbuf.c int strbuf_reencode(struct strbuf *sb, const char *from, const char *to) · line 171
Quoted 10 lines
strbuf.c +1
-1@@ -168,7 +168,7 @@ int strbuf_reencode(struct strbuf *sb, const char *from, const char *to)if (!out)return -1 ; -strbuf_attach (sb, out, len, len); +strbuf_attach (sb, out, len, len +1 );
Not sure that I'm following this change.
t/lib-httpd.sh
Quoted 4 lines
t/lib-httpd.sh +0
-0
I may solicit Peff's input here on the remainder of the test changes, since he is much more familiar with the lib-httpd parts of the suite than I am.
Thanks,
Taylor
Re: [PATCH v2 2/2] http: add trace2 logging for retry operations
In reply to Vaidas Pilkauskas via GitGitGadgetQuoted 3 lines · Vaidas Pilkauskas via GitGitGadget
--- http.c | 23 ++++++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-)
These changes all look reasonable to me. I think you could reaosnably squash this into the previous commit, especially since that commit will likely shrink as you move some hunks out into preparatory patches.
Thanks,
Taylor
Quoted 19 lines · Taylor Blau
Quoted 19 lines
http-push.c +4
-0@@ -716,6 +716,10 @@ static int fetch_indices(void)case HTTP_MISSING_TARGET : ret =0 ;break ; +case HTTP_RATE_LIMITED : +error (_ ("rate limited by '%s', please try again later" ), repo->url); + ret =-1 ; +break ;default : ret =-1 ; }@@ -1548,6 +1552,10 @@ static int remote_exists(const char *path)case HTTP_MISSING_TARGET : ret =0 ;break ;
Quoted 8 lines
http-push.c static int remote_exists(const char *path) · line 1558
Quoted 4 lines
http-push.c +4
-0+case HTTP_RATE_LIMITED : +error (_ ("rate limited by '%s', please try again later" ), url); + ret =-1 ; +break ;I wonder if there is an opportunity to DRY this up a bit? I think the case in
fetch_indices()is very similar toremote_Exists(), and ditto forfetch_indices()in thehttp-walker.ccode.
IMHO it is not worth trying to clean up http-push here. It's the dumb push-over-webdav implementation that nobody uses. I'd actually be happy to see it ripped out, but am too lazy to go through the effort of a big deprecation period myself.
So I would actually consider not touching this code at all, and letting it continue to behave as it did before (returning -1 and not producing any specialized message). Though I suppose in remote_exists() we'd fail to even print the curl error anymore, which would be a regression.
Ditto for http-walker.c's fetch_indices() function. It is used only for dumb-http fetches (which are forbidden by most forges). And if not touched at all, it would continue to function in the same way (not producing any specialized message).
Quoted 4 lines
As a separate but related note, I don't know if this function properly handles header continuations for Retry-After headers, but in practice I suspect it doesn't matter, as servers should not be continuing Retry-After headers across multiple lines.
Yeah, I noticed that, too. And all of the parsing actually makes me nervous. Surely curl can do some of this for us?
...studies some manpages...
Ah, indeed. How about:
curl_off_t wait = 0; curl_easy_getinfo(slot->curl, CURLINFO_RETRY_AFTER, &wait);
You can see how we already dig out similar info in finish_active_slot(). And more extended (but optional) info in http_request(). It looks like CURLINFO_RETRY_AFTER was added in 7.66.0, so this would have to be a conditional feature at build-time. But that seems like a reasonable trade-off.
Side note: the obvious question is why we need fwrite_wwwauth() in the first place. And the answer is that curl does not provide structured access to the information from those headers. It does make me wonder if we could be using curl_easy_header() to get rid of all of this manual parsing and continuation code. That was introduced in 7.83.0, which would again make it conditional. But it seems like a nicer path forward for us. Anyway, way out of scope for this patch.
Quoted 16 lines
Quoted 1 line
http-push.c +0
-0@@ -1660,44 +1729,98 @@ void run_active_slot(struct active_request_slot *slot)[...]
Quoted 14 lines
http-push.c +13
-1-while (!finished) { +while (waiting_for_delay || !finished) { +if (waiting_for_delay) { +gettimeofday (¤t_time,NULL ); + elapsed_time.tv_sec = current_time.tv_sec - start_time.tv_sec; + elapsed_time.tv_usec = current_time.tv_usec - start_time.tv_usec; +if (elapsed_time.tv_usec <0 ) { + elapsed_time.tv_sec--; + elapsed_time.tv_usec +=1000000 ; + } + +if (elapsed_time.tv_sec >= slot->retry_delay_seconds) { + slot->retry_delay_seconds =-1 ; + waiting_for_delay =0 ;
Quoted 8 lines
http-push.c
Quoted 3 lines
http-push.c +3
-0+ +if (slot_not_started) +return ;I wonder if
run_active_slot()is the right place for these changes or if it should be handled separately. I think it may be somewhat surprising forrun_active_slot()to return without actually running the slot, even if the slot is marked as "active" but just waiting for a delay.
Yeah, I agree. The point of run_active_slot() is to run the slot to completion (I think; it has been a while since I've had to dig into any of this). So I'd either expect it to handle the retry and delay itself internally, or to return the failed request to the caller, who will then delay and initiate the retry.
That's all assuming we're making one request at a time (which I think is mostly all that run_active_slot() handles). There's a much more complicated question when we have multiple simultaneous requests, which we'd do only with the dumb protocol (trying to fetch multiple objects at once). In that case we need to be queuing requests. And I _think_ that might be what this code is trying to do. But I'm not sure if it would actually work, as we try to advance those via step_active_slots().
Quoted 4 lines
OTOH, like I mentioned earlier, I am far from an expert in this part of the code, so perhaps this is totally OK. shortlog says that Peff (CC'd) is among the most active contributors to this file in the past year, so I'll be curious what he thinks as well.
Most of the details of this active slot stuff have long been paged out of my memory. It's all _so_ messy because of the desire for the dumb-http code to handle multiple requests. But for smart-http (and I would be perfectly content for this feature to only apply there), we could probably just focus on run_one_slot(), I'd think.
I.e., what I'd expect the simplest form of the patch to look like is roughly:
- teach handle_curl_result() to recognize 429 and pull out the retry-after value, returning HTTP_RETRY
- in run_one_slot(), recognize HTTP_RETRY and if appropriate, sleep and retry
I do wonder if even that might be too low-level, though. For a real large request, we'll be streaming data into the request, and I'm not sure we _can_ retry. We send a probe_rpc() first in that case to try to resolve issues like credential-filling. But there's nothing to say that we can't get a 200 on the probe and a 429 on the real request.
Which I guess implies to me that http_request_reauth() should be where the magic happens. And it somewhat does in this patch, but...why not do the sleeping there, and why push it all the way down into run_active_slot()?
I know I'm kind of talking in circles here, which is indicative of my confusion (and the general complexity of the http code). But as the patch stands, I'm not really convinced which cases it is trying to cover (single requests vs multi, repeatable requests vs streaming POSTs), how well it covers them, and that it is doing it as simply as possible (or at least keeping the logic together).
Quoted 8 lines
Quoted 8 lines
http-push.c +1
-3@@ -518,21 +529,25 @@ static struct discovery *discover_refs(const char *service, int for_push)case HTTP_OK :break ;case HTTP_MISSING_TARGET : -show_http_message (&type, &charset, &buffer); -die (_ ("repository '%s' not found" ), -transport_anonymize_url (url.buf)); + show_http_message_fatal(&type, &charset, &buffer,
Quoted 8 lines
http-push.c static struct discovery *discover_refs(const char *service, int for_push) · line 534
Quoted 2 lines
http-push.c +2
-0+_ ("repository '%s' not found" ), + transport_anonymize_url(url.buf));Thanks for taking my suggestion here as well. I think that the end result reads much cleaner, though I do think that introducing the new
show_http_message_fatal()function and rewriting the existing code should happen in a preparatory commit before this one to more clearly separate the changes.
Yeah, I had the same thought.
Quoted 8 lines
t/lib-httpd.sh
Quoted 4 lines
t/lib-httpd.sh +0
-0I may solicit Peff's input here on the remainder of the test changes, since he is much more familiar with the lib-httpd parts of the suite than I am.
The lib-httpd parts looked about as I'd expect (and I found the use of custom URL components to encode the retry parameters quite clever).
There were lots of uses of "date" that I suspect may give us portability problems. "+%s" is not even in POSIX, but maybe it is universal enough. But stuff like '-d "+2 seconds"' seems likely to be a GNU-ism.
Using "test-tool date" might get around some of that. We even understand relative dates like "2 seconds ago", but I think only in the past. :-/ So you'd probably have to do:
now=$(test-tool date timestamp now | cut -d' ' -f3) then=$((now + 2)) test-tool date show:rfc2822 $then
or something.
-Peff
Quoted 12 lines · Taylor Blau
Documentation/config/http.adoc http.keepAliveCount:: · line 321
Quoted 4 lines
Documentation/config/http.adoc +4
-0+http.retryAfter:: + Default wait time in seconds before retrying when a server returns + HTTP 429 (Too Many Requests) without a Retry-After header. If set + to -1 (the default), Git will fail immediately when encounteringWhile reviewing, I originally wrote:
Setting the default as "-1" makes sense to me. The current behavior is to give up when we receive a HTTP 429 response with or without a Retry-After header, so retaining that behavior makes sense and seems like a sensible path.
Quoted 6 lines
, but I'm not sure that I am sold on that line of thinking. This is controlling how long we'll wait after a 429 response before retrying, not how many times we'll retry (which is
http.maxRetriesbelow).Should the default here be zero? We would "retry" immediately, but that retry would fail since the maximum retries is set to "zero" by default.
I think the only reason I was using "-1" is to have an opportunity to advise on existing configuration for retries, but I guess we can live without advising as I expect folks who are willing to configure retry handling will be advanced users who are aware of the options. I'll switch to "0".
Quoted 5 lines
Quoted 5 lines
http-push.c +0
-0@@ -716,6 +716,10 @@ static int fetch_indices(void)
Quoted 8 lines
http-push.c static int remote_exists(const char *path) · line 1558
Quoted 4 lines
http-push.c +4
-0+case HTTP_RATE_LIMITED : +error (_ ("rate limited by '%s', please try again later" ), url); + ret =-1 ; +break ;I wonder if there is an opportunity to DRY this up a bit? I think the case in
fetch_indices()is very similar toremote_Exists(), and ditto forfetch_indices()in thehttp-walker.ccode.
I'll leave this code unchanged as per Peff's suggestion.
Quoted 24 lines
Quoted 13 lines
http-push.c +12
-0+ slot->results->retry_after = retry_after; + }else { +/* Try parsing as HTTP-date format */ +timestamp_t timestamp; +int offset; +if (!parse_date_basic (buf.buf, ×tamp, &offset)) { +/* Successfully parsed as date, calculate delay from now */ +timestamp_t now =time (NULL ); +if (timestamp > now) { + slot->results->retry_after = (long )(timestamp - now); + }else { +/* Past date means retry immediately */ http-push.c static int fetch_indices(void)
Quoted 10 lines
http-push.c +10
-0+ slot->results->retry_after =0 ; + } + }else { +/* Failed to parse as either delay-seconds or HTTP-date */ +warning (_ ("unable to parse Retry-After header value: '%s'" ), buf.buf); + } + } + } + + http_auth.header_is_last_match =1 ;
Quoted 3 lines
Could you help me understand why we're setting
header_is_last_matchhere? I think since we immediately "goto exit" this line isn't strictly necessary.
Yes, this should not be needed - I'll remove the statement.
Quoted 4 lines
As a separate but related note, I don't know if this function properly handles header continuations for Retry-After headers, but in practice I suspect it doesn't matter, as servers should not be continuing Retry-After headers across multiple lines.
Yes, I assume it's not applicable to Retry-After, so I'm not handling continuations.
Quoted 10 lines
http-push.c void run_active_slot(struct active_request_slot *slot)
Quoted 1 line
http-push.c +0
-0@@ -1660,44 +1729,98 @@ void run_active_slot(struct active_request_slot *slot)I wonder if
run_active_slot()is the right place for these changes or if it should be handled separately. I think it may be somewhat surprising forrun_active_slot()to return without actually running the slot, even if the slot is marked as "active" but just waiting for a delay.OTOH, like I mentioned earlier, I am far from an expert in this part of the code, so perhaps this is totally OK. shortlog says that Peff (CC'd) is among the most active contributors to this file in the past year, so I'll be curious what he thinks as well.
I'll follow Peff's review for this part.
Quoted 10 lines
strbuf.c int strbuf_reencode(struct strbuf *sb, const char *from, const char *to) · line 171
Quoted 10 lines
strbuf.c +1
-1@@ -168,7 +168,7 @@ int strbuf_reencode(struct strbuf *sb, const char *from, const char *to)if (!out)return -1 ; -strbuf_attach (sb, out, len, len); +strbuf_attach (sb, out, len, len +1 );
Sorry, I totally forgot about this change. I still got leak reported from CI, so I
narrowed it down to this line. I'll make a separate commit to discuss it.
Quoted 4 lines
Not sure that I'm following this change.
Thanks,
Taylor
Thanks, Taylor, for the review!
Quoted 7 lines · Jeff King
Yeah, I noticed that, too. And all of the parsing actually makes me nervous. Surely curl can do some of this for us?
...studies some manpages...
Ah, indeed. How about:
Quoted 8 lines
curl_off_t wait = 0; curl_easy_getinfo(slot->curl, CURLINFO_RETRY_AFTER, &wait);You can see how we already dig out similar info in
finish_active_slot(). And more extended (but optional) info inhttp_request(). It looks likeCURLINFO_RETRY_AFTERwas added in 7.66.0, so this would have to be a conditional feature at build-time. But that seems like a reasonable trade-off.
I'll add parsing with libcurl under conditional feature.
Quoted 6 lines
Most of the details of this active slot stuff have long been paged out of my memory. It's all _so_ messy because of the desire for the dumb-http code to handle multiple requests. But for smart-http (and I would be perfectly content for this feature to only apply there), we could probably just focus on
run_one_slot(), I'd think.
Quoted 9 lines
I.e., what I'd expect the simplest form of the patch to look like is roughly:
- teach
handle_curl_result()to recognize 429 and pull out the retry-after value, returningHTTP_RETRY- in
run_one_slot(), recognizeHTTP_RETRYand if appropriate, sleep and retry
This greatly simplifies implementation. I think following similar pattern like auth handling does makes a lot of sense. So, instead of sleeping in run_one_slot(), I think it makes sense to sleep in http_request_recoverable() where HTTP_REAUTH is handled.
Quoted 15 lines
Quoted 3 lines
I may solicit Peff's input here on the remainder of the test changes, since he is much more familiar with the lib-httpd parts of the suite than I am.
The lib-httpd parts looked about as I'd expect (and I found the use of custom URL components to encode the retry parameters quite clever).
There were lots of uses of "date" that I suspect may give us portability problems. "+%s" is not even in POSIX, but maybe it is universal enough. But stuff like '-d "+2 seconds"' seems likely to be a GNU-ism.
Using "test-tool date" might get around some of that. We even understand relative dates like "2 seconds ago", but I think only in the past. :-/ So you'd probably have to do:
Quoted 5 lines
now=$(test-tool date timestamp now | cut -d' ' -f3) then=$((now + 2)) test-tool date show:rfc2822 $thenor something.
I was not aware about test-tool, thanks!
Quoted 1 line
-Peff
Thanks, Peff, for the review!
Quoted 8 lines · Vaidas Pilkauskas
Quoted 8 lines
There were lots of uses of "date" that I suspect may give us portability problems. "+%s" is not even in POSIX, but maybe it is universal enough. But stuff like '-d "+2 seconds"' seems likely to be a GNU-ism.
Using "test-tool date" might get around some of that. We even understand relative dates like "2 seconds ago", but I think only in the past. :-/ So you'd probably have to do:
Quoted 7 lines
Quoted 5 lines
now=$(test-tool date timestamp now | cut -d' ' -f3) then=$((now + 2)) test-tool date show:rfc2822 $thenor something.
I was not aware about test-tool, thanks!
It might be a little awkward to bend it to your will, especially since it likes to print "input -> output" instead of just the output you want. If it gets too hairy, I wouldn't be opposed to teaching it a new option or even a new command-mode for doing this kind of computed date stuff.
-Peff
v3 cover http: add support for HTTP 429 rate limit retries
In reply to Vaidas Pilkauskas via GitGitGadgetChanges since v2:
* New preparatory patch: Introduced show_http_message_fatal() helper function to reduce code duplication in remote-curl.c (suggested by Taylor Blau)
* Removed specific HTTP_RATE_LIMITED error handling from http-push.c and http-walker.c for the obsolete "dumb" protocol, allowing generic error handling to take over (suggested by Jeff King)
* Added support for CURLINFO_RETRY_AFTER on curl >= 7.66.0, falling back to manual header parsing on older versions
* Simplified retry/delay architecture: replaced complex non-blocking "delayed slot" mechanism with simple blocking sleep() call in the retry loop, removing ~66 lines of timing logic (suggested by Jeff King)
* Fixed Retry-After: 0 handling to allow immediate retry as specified by RFC 9110
* Changed http.retryAfter default from -1 to 0, so Git will retry immediately when encountering HTTP 429 without a Retry-After header, rather than failing with a configuration error
* Improved error messages: shortened to be more concise
* Fixed coding style issues: removed unnecessary curly braces, changed x == 0 to !x (per CodingGuidelines)
* Improved test portability: replaced non-portable date(1) commands with test-tool date, added nanosecond-precision timing with getnanos, replaced cut(1) with POSIX shell parameter expansion
* Split out strbuf.c bugfix into separate preparatory patch (the strbuf_reencode alloc size fix is unrelated to HTTP 429 support)
* Squashed separate trace2 logging patch into main HTTP 429 retry support commit
* Kept header_is_last_match assignment for Retry-After to prevent incorrect handling of HTTP header continuation lines
The implementation includes:
1. A bug fix in strbuf_reencode() that corrects the allocation size passed to strbuf_attach(), ensuring proper memory management.
2. A refactoring in remote-curl.c that introduces a show_http_message_fatal() helper to reduce code duplication when handling fatal HTTP errors.
3. The main feature: HTTP 429 retry logic with support for the Retry-After header (both delay-seconds and HTTP-date formats), configurable via http.maxRetries, http.retryAfter, and http.maxRetryTime options.
Vaidas Pilkauskas (3):
strbuf: fix incorrect alloc size in strbuf_reencode()
remote-curl: introduce show_http_message_fatal() helper
http: add support for HTTP 429 rate limit retries
Documentation/config/http.adoc | 23 +++ git-curl-compat.h | 8 + http.c | 190 +++++++++++++++++++++-- http.h | 2 + remote-curl.c | 49 +++--- strbuf.c | 2 +- t/lib-httpd.sh | 1 + t/lib-httpd/apache.conf | 8 + t/lib-httpd/http-429.sh | 98 ++++++++++++ t/meson.build | 1 + t/t5584-http-429-retry.sh | 266 +++++++++++++++++++++++++++++++++ 11 files changed, 618 insertions(+), 30 deletions(-) create mode 100644 t/lib-httpd/http-429.sh create mode 100755 t/t5584-http-429-retry.sh
base-commit: 852829b3dd2fe4e7c7fc4d8badde644cf1b66c74
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-2008%2Fvaidas-shopify%2Fretry-after-v3
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-2008/vaidas-shopify/retry-after-v3
Pull-Request: https://github.com/gitgitgadget/git/pull/2008
Range-diff vs v2:
-: ---------- > 1: 821043c664 strbuf: fix incorrect alloc size in strbuf_reencode()
-: ---------- > 2: 3653067f0e remote-curl: introduce show_http_message_fatal() helper
1: d80ce07703 ! 3: 3cece62a63 http: add support for HTTP 429 rate limit retries
@@ Documentation/config/http.adoc: http.keepAliveCount::
+http.retryAfter::
+ Default wait time in seconds before retrying when a server returns
-+ HTTP 429 (Too Many Requests) without a Retry-After header. If set
-+ to -1 (the default), Git will fail immediately when encountering
-+ a 429 response without a Retry-After header. When a Retry-After
-+ header is present, its value takes precedence over this setting.
-+ Can be overridden by the `GIT_HTTP_RETRY_AFTER` environment variable.
++ HTTP 429 (Too Many Requests) without a Retry-After header.
++ Defaults to 0 (retry immediately). When a Retry-After header is
++ present, its value takes precedence over this setting. Can be
++ overridden by the `GIT_HTTP_RETRY_AFTER` environment variable.
+ See also `http.maxRetries` and `http.maxRetryTime`.
+
+http.maxRetries::
@@ Documentation/config/http.adoc: http.keepAliveCount::
A boolean which disables using of EPSV ftp command by curl.
This can be helpful with some "poor" ftp servers which don't
- ## http-push.c ##
-@@ http-push.c: static int fetch_indices(void)
- case HTTP_MISSING_TARGET:
- ret = 0;
- break;
-+ case HTTP_RATE_LIMITED:
-+ error(_("rate limited by '%s', please try again later"), repo->url);
-+ ret = -1;
-+ break;
- default:
- ret = -1;
- }
-@@ http-push.c: static int remote_exists(const char *path)
- case HTTP_MISSING_TARGET:
- ret = 0;
- break;
-+ case HTTP_RATE_LIMITED:
-+ error(_("rate limited by '%s', please try again later"), url);
-+ ret = -1;
-+ break;
- case HTTP_ERROR:
- error("unable to access '%s': %s", url, curl_errorstr);
- /* fallthrough */
-
- ## http-walker.c ##
-@@ http-walker.c: static int fetch_indices(struct walker *walker, struct alt_base *repo)
- repo->got_indices = 1;
- ret = 0;
- break;
-+ case HTTP_RATE_LIMITED:
-+ error("rate limited by '%s', please try again later", repo->base);
-+ repo->got_indices = 0;
-+ ret = -1;
-+ break;
- default:
- repo->got_indices = 0;
- ret = -1;
+ ## git-curl-compat.h ##
+@@
+ #define GIT_CURL_NEED_TRANSFER_ENCODING_HEADER
+ #endif
+
++/**
++ * CURLINFO_RETRY_AFTER was added in 7.66.0, released in September 2019.
++ * It allows curl to automatically parse Retry-After headers.
++ */
++#if LIBCURL_VERSION_NUM >= 0x074200
++#define GIT_CURL_HAVE_CURLINFO_RETRY_AFTER 1
++#endif
++
+ /**
+ * CURLOPT_PROTOCOLS_STR and CURLOPT_REDIR_PROTOCOLS_STR were added in 7.85.0,
+ * released in August 2022.
## http.c ##
@@
@@ http.c
#include "odb.h"
#include "tempfile.h"
+#include "date.h"
++#include "trace2.h"
static struct trace_key trace_curl = TRACE_KEY_INIT(CURL);
static int trace_curl_data = 1;
@@ http.c: static char *cached_accept_language;
static int http_schannel_check_revoke = 1;
+
-+static long http_retry_after = -1;
++static long http_retry_after = 0;
+static long http_max_retries = 0;
+static long http_max_retry_time = 300;
+
@@ http.c: static inline int is_hdr_continuation(const char *ptr, const size_t size
}
-static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UNUSED)
-+static size_t fwrite_headers(char *ptr, size_t eltsize, size_t nmemb, void *p)
++static size_t fwrite_headers(char *ptr, size_t eltsize, size_t nmemb, void *p MAYBE_UNUSED)
{
size_t size = eltsize * nmemb;
struct strvec *values = &http_auth.wwwauth_headers;
- struct strbuf buf = STRBUF_INIT;
- const char *val;
- size_t val_len;
-+ struct active_request_slot *slot = (struct active_request_slot *)p;
-
- /*
- * Header lines may not come NULL-terminated from libcurl so we must
@@ http.c: static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UN
goto exit;
}
-+ /* Parse Retry-After header for rate limiting */
++#ifndef GIT_CURL_HAVE_CURLINFO_RETRY_AFTER
++ /* Parse Retry-After header for rate limiting (for curl < 7.66.0) */
+ if (skip_iprefix_mem(ptr, size, "retry-after:", &val, &val_len)) {
++ struct active_request_slot *slot = (struct active_request_slot *)p;
++
+ strbuf_add(&buf, val, val_len);
+ strbuf_trim(&buf);
+
@@ http.c: static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, vo
+ errno = 0;
+ retry_after = strtol(buf.buf, &endptr, 10);
+
-+ /* Check if it's a valid integer (delay-seconds format) */
-+ if (endptr != buf.buf && *endptr == '\0' &&
-+ errno != ERANGE && retry_after > 0) {
-+ slot->results->retry_after = retry_after;
-+ } else {
++ /* Check if it's a valid integer (delay-seconds format) */
++ if (endptr != buf.buf && *endptr == '\0' &&
++ errno != ERANGE && retry_after >= 0) {
++ slot->results->retry_after = retry_after;
++ } else {
+ /* Try parsing as HTTP-date format */
+ timestamp_t timestamp;
+ int offset;
@@ http.c: static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, vo
+ }
+ }
+
-+ http_auth.header_is_last_match = 1;
+ goto exit;
+ }
++#endif
+
/*
* This line could be a continuation of the previously matched header
* field. If this is the case then we should append this value to the
+@@ http.c: static void finish_active_slot(struct active_request_slot *slot)
+
+ curl_easy_getinfo(slot->curl, CURLINFO_HTTP_CONNECTCODE,
+ &slot->results->http_connectcode);
++
++#ifdef GIT_CURL_HAVE_CURLINFO_RETRY_AFTER
++ if (slot->results->http_code == 429) {
++ curl_off_t retry_after;
++ CURLcode res = curl_easy_getinfo(slot->curl,
++ CURLINFO_RETRY_AFTER,
++ &retry_after);
++ if (res == CURLE_OK && retry_after > 0)
++ slot->results->retry_after = (long)retry_after;
++ }
++#endif
+ }
+
+ /* Run callback if appropriate */
@@ http.c: static int http_options(const char *var, const char *value,
return 0;
}
@@ http.c: void http_init(struct remote *remote, const char *url, int proactive_aut
curl_default = get_curl_handle();
}
-@@ http.c: struct active_request_slot *get_active_slot(void)
- slot->finished = NULL;
- slot->callback_data = NULL;
- slot->callback_func = NULL;
-+ slot->retry_delay_seconds = -1;
-+ memset(&slot->retry_delay_start, 0, sizeof(slot->retry_delay_start));
-
- if (curl_cookie_file && !strcmp(curl_cookie_file, "-")) {
- warning(_("refusing to read cookies from http.cookiefile '-'"));
-@@ http.c: void run_active_slot(struct active_request_slot *slot)
- fd_set excfds;
- int max_fd;
- struct timeval select_timeout;
-+ long curl_timeout;
-+ struct timeval start_time = {0}, current_time, elapsed_time = {0};
-+ long remaining_seconds;
- int finished = 0;
-+ int slot_not_started = (slot->finished == NULL);
-+ int waiting_for_delay = (slot->retry_delay_seconds > 0);
-+
-+ if (waiting_for_delay) {
-+ warning(_("rate limited, waiting %ld seconds before retry"), slot->retry_delay_seconds);
-+ start_time = slot->retry_delay_start;
-+ }
-
- slot->finished = &finished;
-- while (!finished) {
-+ while (waiting_for_delay || !finished) {
-+ if (waiting_for_delay) {
-+ gettimeofday(¤t_time, NULL);
-+ elapsed_time.tv_sec = current_time.tv_sec - start_time.tv_sec;
-+ elapsed_time.tv_usec = current_time.tv_usec - start_time.tv_usec;
-+ if (elapsed_time.tv_usec < 0) {
-+ elapsed_time.tv_sec--;
-+ elapsed_time.tv_usec += 1000000;
-+ }
-+
-+ if (elapsed_time.tv_sec >= slot->retry_delay_seconds) {
-+ slot->retry_delay_seconds = -1;
-+ waiting_for_delay = 0;
-+
-+ if (slot_not_started)
-+ return;
-+ }
-+ }
-+
- step_active_slots();
-
-- if (slot->in_use) {
-- long curl_timeout;
-- curl_multi_timeout(curlm, &curl_timeout);
-- if (curl_timeout == 0) {
-+ if (!waiting_for_delay && !slot->in_use)
-+ continue;
-+
-+ curl_multi_timeout(curlm, &curl_timeout);
-+ if (curl_timeout == 0) {
-+ if (!waiting_for_delay)
- continue;
-- } else if (curl_timeout == -1) {
-- select_timeout.tv_sec = 0;
-- select_timeout.tv_usec = 50000;
-+ select_timeout.tv_sec = 0;
-+ select_timeout.tv_usec = 50000; /* 50ms */
-+ } else if (curl_timeout == -1) {
-+ select_timeout.tv_sec = 0;
-+ select_timeout.tv_usec = 50000;
-+ } else {
-+ long curl_timeout_sec = curl_timeout / 1000;
-+ long curl_timeout_usec = (curl_timeout % 1000) * 1000;
-+
-+ if (waiting_for_delay) {
-+ remaining_seconds = slot->retry_delay_seconds - elapsed_time.tv_sec;
-+ if (curl_timeout_sec < remaining_seconds) {
-+ select_timeout.tv_sec = curl_timeout_sec;
-+ select_timeout.tv_usec = curl_timeout_usec;
-+ } else {
-+ select_timeout.tv_sec = remaining_seconds;
-+ select_timeout.tv_usec = 0;
-+ }
- } else {
-- select_timeout.tv_sec = curl_timeout / 1000;
-- select_timeout.tv_usec = (curl_timeout % 1000) * 1000;
-+ select_timeout.tv_sec = curl_timeout_sec;
-+ select_timeout.tv_usec = curl_timeout_usec;
- }
-+ }
-
-- max_fd = -1;
-- FD_ZERO(&readfds);
-- FD_ZERO(&writefds);
-- FD_ZERO(&excfds);
-- curl_multi_fdset(curlm, &readfds, &writefds, &excfds, &max_fd);
-+ max_fd = -1;
-+ FD_ZERO(&readfds);
-+ FD_ZERO(&writefds);
-+ FD_ZERO(&excfds);
-+ curl_multi_fdset(curlm, &readfds, &writefds, &excfds, &max_fd);
-
-- /*
-- * It can happen that curl_multi_timeout returns a pathologically
-- * long timeout when curl_multi_fdset returns no file descriptors
-- * to read. See commit message for more details.
-- */
-- if (max_fd < 0 &&
-- (select_timeout.tv_sec > 0 ||
-- select_timeout.tv_usec > 50000)) {
-- select_timeout.tv_sec = 0;
-- select_timeout.tv_usec = 50000;
-- }
-+ /*
-+ * It can happen that curl_multi_timeout returns a pathologically
-+ * long timeout when curl_multi_fdset returns no file descriptors
-+ * to read. See commit message for more details.
-+ */
-+ if (max_fd < 0 &&
-+ (select_timeout.tv_sec > 0 ||
-+ select_timeout.tv_usec > 50000)) {
-+ select_timeout.tv_sec = 0;
-+ select_timeout.tv_usec = 50000;
-+ }
-
-- select(max_fd+1, &readfds, &writefds, &excfds, &select_timeout);
-+ /*
-+ * If curl_multi_fdset returns no file descriptors but we have
-+ * a timeout, still use select() to wait for the timeout period.
-+ */
-+ if (max_fd < 0) {
-+ /* No file descriptors, just wait for timeout */
-+ select(0, NULL, NULL, NULL, &select_timeout);
-+ } else {
-+ select(max_fd + 1, &readfds, &writefds, &excfds, &select_timeout);
- }
- }
-
@@ http.c: static int handle_curl_result(struct slot_results *results)
}
return HTTP_REAUTH;
}
+ } else if (results->http_code == 429) {
++ trace2_data_intmax("http", the_repository, "http/429-retry-after",
++ results->retry_after);
+ return HTTP_RATE_LIMITED;
} else {
if (results->http_connectcode == 407)
@@ http.c: int run_one_slot(struct active_request_slot *slot,
slot->results = results;
+ /* Initialize retry_after to -1 (not set) */
+ results->retry_after = -1;
-+
-+ /* If there's a retry delay, wait for it before starting the slot */
-+ if (slot->retry_delay_seconds > 0) {
-+ run_active_slot(slot);
-+ }
+
if (!start_active_slot(slot)) {
xsnprintf(curl_errorstr, sizeof(curl_errorstr),
"failed to start HTTP request");
@@ http.c: static void http_opt_request_remainder(CURL *curl, off_t pos)
- #define HTTP_REQUEST_STRBUF 0
- #define HTTP_REQUEST_FILE 1
-+static void sleep_for_retry(struct active_request_slot *slot, long retry_after);
-+
static int http_request(const char *url,
void *result, int target,
- const struct http_get_options *options)
+ const struct http_get_options *options,
-+ long *retry_after_out,
-+ long retry_delay)
++ long *retry_after_out)
{
struct active_request_slot *slot;
struct slot_results results;
-@@ http.c: static int http_request(const char *url,
- int ret;
-
- slot = get_active_slot();
-+ /* Mark slot for delay if retry delay is provided */
-+ if (retry_delay > 0) {
-+ sleep_for_retry(slot, retry_delay);
-+ }
- curl_easy_setopt(slot->curl, CURLOPT_HTTPGET, 1L);
-
- if (!result) {
@@ http.c: static int http_request(const char *url,
fwrite_buffer);
}
@@ http.c: static int update_url_from_redirect(struct strbuf *base,
-static int http_request_reauth(const char *url,
+/*
-+ * Mark slot to be delayed for retry. The actual delay will be handled
-+ * in run_active_slot when the slot is executed.
-+ */
-+static void sleep_for_retry(struct active_request_slot *slot, long retry_after)
-+{
-+ if (retry_after > 0 && slot) {
-+ warning(_("rate limited, waiting %ld seconds before retry"), retry_after);
-+ slot->retry_delay_seconds = retry_after;
-+ gettimeofday(&slot->retry_delay_start, NULL);
-+ }
-+}
-+
-+/*
+ * Handle rate limiting retry logic for HTTP 429 responses.
-+ * Uses slot-specific retry_after value to support concurrent slots.
+ * Returns a negative value if retries are exhausted or configuration is invalid,
+ * otherwise returns the delay value (>= 0) to indicate the retry should proceed.
+ */
+static long handle_rate_limit_retry(int *rate_limit_retries, long slot_retry_after)
+{
+ int retry_attempt = http_max_retries - *rate_limit_retries + 1;
++
++ trace2_data_intmax("http", the_repository, "http/429-retry-attempt",
++ retry_attempt);
++
+ if (*rate_limit_retries <= 0) {
+ /* Retries are disabled or exhausted */
+ if (http_max_retries > 0) {
+ error(_("too many rate limit retries, giving up"));
++ trace2_data_string("http", the_repository,
++ "http/429-error", "retries-exhausted");
+ }
+ return -1;
+ }
+
-+ /* Decrement retries counter */
+ (*rate_limit_retries)--;
+
+ /* Use the slot-specific retry_after value or configured default */
+ if (slot_retry_after >= 0) {
+ /* Check if retry delay exceeds maximum allowed */
+ if (slot_retry_after > http_max_retry_time) {
-+ error(_("rate limited (HTTP 429) requested %ld second delay, "
-+ "exceeds http.maxRetryTime of %ld seconds"),
++ error(_("response requested a delay greater than http.maxRetryTime (%ld > %ld seconds)"),
+ slot_retry_after, http_max_retry_time);
++ trace2_data_string("http", the_repository,
++ "http/429-error", "exceeds-max-retry-time");
++ trace2_data_intmax("http", the_repository,
++ "http/429-requested-delay", slot_retry_after);
+ return -1;
+ }
+ return slot_retry_after;
+ } else {
-+ /* No Retry-After header provided */
-+ if (http_retry_after < 0) {
-+ /* Not configured - exit with error */
-+ error(_("rate limited (HTTP 429) and no Retry-After header provided. "
-+ "Configure http.retryAfter or set GIT_HTTP_RETRY_AFTER."));
-+ return -1;
-+ }
-+ /* Check if configured default exceeds maximum allowed */
++ /* No Retry-After header provided, use configured default */
+ if (http_retry_after > http_max_retry_time) {
-+ error(_("configured http.retryAfter (%ld seconds) exceeds "
-+ "http.maxRetryTime (%ld seconds)"),
++ error(_("configured http.retryAfter exceeds http.maxRetryTime (%ld > %ld seconds)"),
+ http_retry_after, http_max_retry_time);
++ trace2_data_string("http", the_repository,
++ "http/429-error", "config-exceeds-max-retry-time");
+ return -1;
+ }
-+
++ trace2_data_string("http", the_repository,
++ "http/429-retry-source", "config-default");
+ return http_retry_after;
+ }
+}
@@ http.c: static int update_url_from_redirect(struct strbuf *base,
credential_fill(the_repository, &http_auth, 1);
- ret = http_request(url, result, target, options);
-+ ret = http_request(url, result, target, options, &slot_retry_after, -1);
++ ret = http_request(url, result, target, options, &slot_retry_after);
- if (ret != HTTP_OK && ret != HTTP_REAUTH)
+ if (ret != HTTP_OK && ret != HTTP_REAUTH && ret != HTTP_RATE_LIMITED)
return ret;
+ /* If retries are disabled and we got a 429, fail immediately */
-+ if (ret == HTTP_RATE_LIMITED && http_max_retries == 0)
++ if (ret == HTTP_RATE_LIMITED && !http_max_retries)
+ return HTTP_ERROR;
+
if (options && options->effective_url && options->base_url) {
@@ http.c: static int http_request_reauth(const char *url,
+ retry_delay = handle_rate_limit_retry(&rate_limit_retries, slot_retry_after);
+ if (retry_delay < 0)
+ return HTTP_ERROR;
++
++ if (retry_delay > 0) {
++ warning(_("rate limited, waiting %ld seconds before retry"), retry_delay);
++ trace2_data_intmax("http", the_repository,
++ "http/retry-sleep-seconds", retry_delay);
++ sleep(retry_delay);
++ }
+ slot_retry_after = -1; /* Reset after use */
+ } else if (ret == HTTP_REAUTH) {
+ credential_fill(the_repository, &http_auth, 1);
@@ http.c: static int http_request_reauth(const char *url,
- credential_fill(the_repository, &http_auth, 1);
-
- ret = http_request(url, result, target, options);
-+ ret = http_request(url, result, target, options, &slot_retry_after, retry_delay);
++ ret = http_request(url, result, target, options, &slot_retry_after);
}
return ret;
}
@@ http.h: struct slot_results {
};
struct active_request_slot {
-@@ http.h: struct active_request_slot {
- void *callback_data;
- void (*callback_func)(void *data);
- struct active_request_slot *next;
-+ long retry_delay_seconds;
-+ struct timeval retry_delay_start;
- };
-
- struct buffer {
@@ http.h: struct http_get_options {
#define HTTP_REAUTH 4
#define HTTP_NOAUTH 5
@@ http.h: struct http_get_options {
* Requests a URL and stores the result in a strbuf.
## remote-curl.c ##
-@@ remote-curl.c: static void free_discovery(struct discovery *d)
- }
- }
-
--static int show_http_message(struct strbuf *type, struct strbuf *charset,
-- struct strbuf *msg)
-+static NORETURN void show_http_message_fatal(struct strbuf *type, struct strbuf *charset,
-+ struct strbuf *msg, const char *fmt, ...)
- {
- const char *p, *eol;
-+ va_list ap;
-+ report_fn die_message_routine = get_die_message_routine();
-
- /*
- * We only show text/plain parts, as other types are likely
- * to be ugly to look at on the user's terminal.
- */
- if (strcmp(type->buf, "text/plain"))
-- return -1;
-+ goto out;
- if (charset->len)
- strbuf_reencode(msg, charset->buf, get_log_output_encoding());
-
- strbuf_trim(msg);
- if (!msg->len)
-- return -1;
-+ goto out;
-
- p = msg->buf;
- do {
-@@ remote-curl.c: static int show_http_message(struct strbuf *type, struct strbuf *charset,
- fprintf(stderr, "remote: %.*s\n", (int)(eol - p), p);
- p = eol + 1;
- } while(*eol);
-- return 0;
-+
-+out:
-+ strbuf_release(type);
-+ strbuf_release(charset);
-+ strbuf_release(msg);
-+
-+ va_start(ap, fmt);
-+ die_message_routine(fmt, ap);
-+ va_end(ap);
-+ exit(128);
- }
-
- static int get_protocol_http_header(enum protocol_version version,
@@ remote-curl.c: static struct discovery *discover_refs(const char *service, int for_push)
- case HTTP_OK:
- break;
- case HTTP_MISSING_TARGET:
-- show_http_message(&type, &charset, &buffer);
-- die(_("repository '%s' not found"),
-- transport_anonymize_url(url.buf));
-+ show_http_message_fatal(&type, &charset, &buffer,
-+ _("repository '%s' not found"),
-+ transport_anonymize_url(url.buf));
- case HTTP_NOAUTH:
-- show_http_message(&type, &charset, &buffer);
-- die(_("Authentication failed for '%s'"),
-- transport_anonymize_url(url.buf));
-+ show_http_message_fatal(&type, &charset, &buffer,
-+ _("Authentication failed for '%s'"),
-+ transport_anonymize_url(url.buf));
- case HTTP_NOMATCHPUBLICKEY:
-- show_http_message(&type, &charset, &buffer);
-- die(_("unable to access '%s' with http.pinnedPubkey configuration: %s"),
-- transport_anonymize_url(url.buf), curl_errorstr);
-+ show_http_message_fatal(&type, &charset, &buffer,
-+ _("unable to access '%s' with http.pinnedPubkey configuration: %s"),
-+ transport_anonymize_url(url.buf), curl_errorstr);
+ show_http_message_fatal(&type, &charset, &buffer,
+ _("unable to access '%s' with http.pinnedPubkey configuration: %s"),
+ transport_anonymize_url(url.buf), curl_errorstr);
+ case HTTP_RATE_LIMITED:
+ show_http_message_fatal(&type, &charset, &buffer,
+ _("rate limited by '%s', please try again later"),
+ transport_anonymize_url(url.buf));
default:
-- show_http_message(&type, &charset, &buffer);
-- die(_("unable to access '%s': %s"),
-- transport_anonymize_url(url.buf), curl_errorstr);
-+ show_http_message_fatal(&type, &charset, &buffer,
-+ _("unable to access '%s': %s"),
-+ transport_anonymize_url(url.buf), curl_errorstr);
- }
-
- if (options.verbosity && !starts_with(refs_url.buf, url.buf)) {
-
- ## strbuf.c ##
-@@ strbuf.c: int strbuf_reencode(struct strbuf *sb, const char *from, const char *to)
- if (!out)
- return -1;
-
-- strbuf_attach(sb, out, len, len);
-+ strbuf_attach(sb, out, len, len + 1);
- return 0;
- }
-
+ show_http_message_fatal(&type, &charset, &buffer,
+ _("unable to access '%s': %s"),
## t/lib-httpd.sh ##
@@ t/lib-httpd.sh: prepare_httpd() {
@@ t/t5584-http-429-retry.sh (new)
+ test_config http.maxRetries 3 &&
+
+ # Time the operation - it should take at least 2 seconds due to retry delay
-+ start=$(date +%s) &&
++ start=$(test-tool date getnanos) &&
+ git ls-remote "$HTTPD_URL/http_429/retry-delays-respected/2/repo.git" >output 2>err &&
-+ end=$(date +%s) &&
-+ duration=$((end - start)) &&
++ duration=$(test-tool date getnanos $start) &&
+
+ # Verify it took at least 2 seconds (allowing some tolerance)
-+ test "$duration" -ge 1 &&
++ duration_int=${duration%.*} &&
++ test "$duration_int" -ge 1 &&
+ test_grep "refs/heads/" output
+'
+
@@ t/t5584-http-429-retry.sh (new)
+ test_config http.maxRetryTime 3 &&
+
+ # Should fail immediately without waiting
-+ start=$(date +%s) &&
++ start=$(test-tool date getnanos) &&
+ test_must_fail git ls-remote "$HTTPD_URL/http_429/retry-after-exceeds-max-time/100/repo.git" 2>err &&
-+ end=$(date +%s) &&
-+ duration=$((end - start)) &&
++ duration=$(test-tool date getnanos $start) &&
+
+ # Should fail quickly (less than 2 seconds, no 100 second wait)
-+ test "$duration" -lt 2 &&
-+ test_grep "exceeds http.maxRetryTime" err
++ duration_int=${duration%.*} &&
++ test "$duration_int" -lt 2 &&
++ test_grep "greater than http.maxRetryTime" err
+'
+
+test_expect_success 'HTTP 429 fails if configured http.retryAfter exceeds http.maxRetryTime' '
@@ t/t5584-http-429-retry.sh (new)
+ test_config http.maxRetryTime 5 &&
+
+ # Should fail immediately with configuration error
-+ start=$(date +%s) &&
++ start=$(test-tool date getnanos) &&
+ test_must_fail git ls-remote "$HTTPD_URL/http_429/config-retry-after-exceeds-max-time/none/repo.git" 2>err &&
-+ end=$(date +%s) &&
-+ duration=$((end - start)) &&
++ duration=$(test-tool date getnanos $start) &&
+
+ # Should fail quickly
-+ test "$duration" -lt 2 &&
++ duration_int=${duration%.*} &&
++ test "$duration_int" -lt 2 &&
+ test_grep "configured http.retryAfter.*exceeds.*http.maxRetryTime" err
+'
+
+test_expect_success 'HTTP 429 with Retry-After HTTP-date format' '
+ # Test HTTP-date format (RFC 2822) in Retry-After header
-+ # Generate a date 2 seconds in the future
-+ future_date=$(TZ=GMT date -d "+2 seconds" "+%a, %d %b %Y %H:%M:%S GMT" 2>/dev/null || \
-+ TZ=GMT date -v+2S "+%a, %d %b %Y %H:%M:%S GMT" 2>/dev/null || \
-+ echo "skip") &&
-+
-+ if test "$future_date" = "skip"
-+ then
-+ skip_all="date command does not support required format" &&
-+ test_done
-+ fi &&
-+
-+ # URL-encode the date (replace spaces with %20)
++ raw=$(test-tool date timestamp now) &&
++ now="${raw#* -> }" &&
++ future_time=$((now + 2)) &&
++ raw=$(test-tool date show:rfc2822 $future_time) &&
++ future_date="${raw#* -> }" &&
+ future_date_encoded=$(echo "$future_date" | sed "s/ /%20/g") &&
+
+ # Enable retries
+ test_config http.maxRetries 3 &&
+
+ # Git should parse the HTTP-date and retry after the delay
-+ start=$(date +%s) &&
++ start=$(test-tool date getnanos) &&
+ git ls-remote "$HTTPD_URL/http_429/http-date-format/$future_date_encoded/repo.git" >output 2>err &&
-+ end=$(date +%s) &&
-+ duration=$((end - start)) &&
++ duration=$(test-tool date getnanos $start) &&
+
+ # Should take at least 1 second (allowing tolerance for processing time)
-+ test "$duration" -ge 1 &&
++ duration_int=${duration%.*} &&
++ test "$duration_int" -ge 1 &&
+ test_grep "refs/heads/" output
+'
+
+test_expect_success 'HTTP 429 with HTTP-date exceeding maxRetryTime fails immediately' '
-+ # Generate a date 200 seconds in the future
-+ future_date=$(TZ=GMT date -d "+200 seconds" "+%a, %d %b %Y %H:%M:%S GMT" 2>/dev/null || \
-+ TZ=GMT date -v+200S "+%a, %d %b %Y %H:%M:%S GMT" 2>/dev/null || \
-+ echo "skip") &&
-+
-+ if test "$future_date" = "skip"
-+ then
-+ skip_all="date command does not support required format" &&
-+ test_done
-+ fi &&
-+
-+ # URL-encode the date (replace spaces with %20)
++ raw=$(test-tool date timestamp now) &&
++ now="${raw#* -> }" &&
++ future_time=$((now + 200)) &&
++ raw=$(test-tool date show:rfc2822 $future_time) &&
++ future_date="${raw#* -> }" &&
+ future_date_encoded=$(echo "$future_date" | sed "s/ /%20/g") &&
+
+ # Configure max retry time much less than the 200 second delay
@@ t/t5584-http-429-retry.sh (new)
+ test_config http.maxRetryTime 10 &&
+
+ # Should fail immediately without waiting 200 seconds
-+ start=$(date +%s) &&
++ start=$(test-tool date getnanos) &&
+ test_must_fail git ls-remote "$HTTPD_URL/http_429/http-date-exceeds-max-time/$future_date_encoded/repo.git" 2>err &&
-+ end=$(date +%s) &&
-+ duration=$((end - start)) &&
++ duration=$(test-tool date getnanos $start) &&
+
+ # Should fail quickly (not wait 200 seconds)
-+ test "$duration" -lt 2 &&
-+ test_grep "exceeds http.maxRetryTime" err
++ duration_int=${duration%.*} &&
++ test "$duration_int" -lt 2 &&
++ test_grep "http.maxRetryTime" err
+'
+
+test_expect_success 'HTTP 429 with past HTTP-date should not wait' '
-+ past_date=$(TZ=GMT date -d "-10 seconds" "+%a, %d %b %Y %H:%M:%S GMT" 2>/dev/null || \
-+ TZ=GMT date -v-10S "+%a, %d %b %Y %H:%M:%S GMT" 2>/dev/null || \
-+ echo "skip") &&
-+
-+ if test "$past_date" = "skip"
-+ then
-+ skip_all="date command does not support required format" &&
-+ test_done
-+ fi &&
-+
-+ # URL-encode the date (replace spaces with %20)
++ raw=$(test-tool date timestamp now) &&
++ now="${raw#* -> }" &&
++ past_time=$((now - 10)) &&
++ raw=$(test-tool date show:rfc2822 $past_time) &&
++ past_date="${raw#* -> }" &&
+ past_date_encoded=$(echo "$past_date" | sed "s/ /%20/g") &&
+
+ # Enable retries
+ test_config http.maxRetries 3 &&
+
+ # Git should retry immediately without waiting
-+ start=$(date +%s) &&
++ start=$(test-tool date getnanos) &&
+ git ls-remote "$HTTPD_URL/http_429/past-http-date/$past_date_encoded/repo.git" >output 2>err &&
-+ end=$(date +%s) &&
-+ duration=$((end - start)) &&
++ duration=$(test-tool date getnanos $start) &&
+
+ # Should complete quickly (less than 2 seconds)
-+ test "$duration" -lt 2 &&
++ duration_int=${duration%.*} &&
++ test "$duration_int" -lt 2 &&
+ test_grep "refs/heads/" output
+'
+
@@ t/t5584-http-429-retry.sh (new)
+ test_config http.retryAfter 1 &&
+
+ # Should use configured default (1 second) since header is invalid
-+ start=$(date +%s) &&
++ start=$(test-tool date getnanos) &&
+ git ls-remote "$HTTPD_URL/http_429/invalid-retry-after-format/invalid/repo.git" >output 2>err &&
-+ end=$(date +%s) &&
-+ duration=$((end - start)) &&
++ duration=$(test-tool date getnanos $start) &&
+
+ # Should take at least 1 second (the configured default)
-+ test "$duration" -ge 1 &&
++ duration_int=${duration%.*} &&
++ test "$duration_int" -ge 1 &&
+ test_grep "refs/heads/" output &&
+ test_grep "waiting.*retry" err
+'
@@ t/t5584-http-429-retry.sh (new)
+ test_config http.retryAfter 10 &&
+
+ # Override with environment variable to 1 second
-+ start=$(date +%s) &&
++ start=$(test-tool date getnanos) &&
+ GIT_HTTP_RETRY_AFTER=1 git ls-remote "$HTTPD_URL/http_429/env-retry-after-override/none/repo.git" >output 2>err &&
-+ end=$(date +%s) &&
-+ duration=$((end - start)) &&
++ duration=$(test-tool date getnanos $start) &&
+
+ # Should use env var (1 second), not config (10 seconds)
-+ test "$duration" -ge 1 &&
-+ test "$duration" -lt 5 &&
++ duration_int=${duration%.*} &&
++ test "$duration_int" -ge 1 &&
++ test "$duration_int" -lt 5 &&
+ test_grep "refs/heads/" output &&
+ test_grep "waiting.*retry" err
+'
@@ t/t5584-http-429-retry.sh (new)
+ test_config http.maxRetryTime 100 &&
+
+ # Override with environment variable to 10 seconds (should reject 50 second delay)
-+ start=$(date +%s) &&
++ start=$(test-tool date getnanos) &&
+ test_must_fail env GIT_HTTP_MAX_RETRY_TIME=10 \
+ git ls-remote "$HTTPD_URL/http_429/env-max-retry-time-override/50/repo.git" 2>err &&
-+ end=$(date +%s) &&
-+ duration=$((end - start)) &&
++ duration=$(test-tool date getnanos $start) &&
+
+ # Should fail quickly (not wait 50 seconds) because env var limits to 10
-+ test "$duration" -lt 5 &&
-+ test_grep "exceeds http.maxRetryTime" err
++ duration_int=${duration%.*} &&
++ test "$duration_int" -lt 5 &&
++ test_grep "greater than http.maxRetryTime" err
+'
+
+test_expect_success 'verify normal repository access still works' ' 2: ad4495fc94 < -: ---------- http: add trace2 logging for retry operations
Signature · 1 line
gitgitgadget
PATCH v3 1/3 strbuf: fix incorrect alloc size in strbuf_reencode()
From: Vaidas Pilkauskas <[email protected]>
The strbuf_reencode() function incorrectly passes the string length as the allocation size to strbuf_attach(), when it should pass length + 1 to account for the null terminator.
The reencode_string_len() function allocates len + 1 bytes (including the null terminator) and returns the string length (excluding the null terminator) via the len parameter. However, strbuf_reencode() then calls strbuf_attach() with this length value as both the len and alloc parameters:
strbuf_attach(sb, out, len, len);
This is incorrect because strbuf_attach()'s alloc parameter should reflect the actual allocated buffer size, which includes space for the null terminator. This could lead to incorrect memory management in code that relies on sb->alloc being accurate.
Fix by passing len + 1 as the alloc parameter:
strbuf_attach(sb, out, len, len + 1);
--- strbuf.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
strbuf.c +1 -1
@@ -168,7 +168,7 @@ int strbuf_reencode(struct strbuf *sb, const char *from, const char *to)if (!out)return -1 ; -strbuf_attach (sb, out, len, len); +strbuf_attach (sb, out, len, len +1 );return 0 ; }
Signature · 1 line
gitgitgadget
PATCH v3 2/3 remote-curl: introduce show_http_message_fatal() helper
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
Several code paths in remote-curl.c follow the same pattern of calling show_http_message() to display server error messages followed by die() to terminate with an error. This duplication makes the code more verbose and harder to maintain.
Introduce a new show_http_message_fatal() helper function that combines these two operations. This function:
1. Displays any HTTP error message from the server via show_http_message()
2. Calls die() with the provided error message
3. Returns NORETURN to help the compiler with control flow analysis
Refactor existing call sites in remote-curl.c to use this new helper, reducing code duplication and improving readability. This pattern will also be used by upcoming HTTP 429 rate limiting support.
--- remote-curl.c | 45 ++++++++++++++++++++++++++++----------------- 1 file changed, 28 insertions(+), 17 deletions(-)
remote-curl.c +28 -17
@@ -367,23 +367,25 @@ static void free_discovery(struct discovery *d) } } -static int show_http_message (struct strbuf *type,struct strbuf *charset, -struct strbuf *msg) +static NORETURN void show_http_message_fatal(struct strbuf *type, struct strbuf *charset, +struct strbuf *msg,const char *fmt, ...) {const char *p, *eol; +va_list ap; +report_fn die_message_routine =get_die_message_routine ();/* * We only show text/plain parts, as other types are likely * to be ugly to look at on the user's terminal. */ if (strcmp (type->buf,"text/plain" )) -return -1 ; + goto out;if (charset->len)strbuf_reencode (msg, charset->buf,get_log_output_encoding ());strbuf_trim (msg);if (!msg->len) -return -1 ; + goto out; p = msg->buf; do {
@@ -391,7 +393,16 @@ static int show_http_message(struct strbuf *type, struct strbuf *charset,fprintf (stderr,"remote: %.*s\n" , (int )(eol - p), p); p = eol +1 ; }while (*eol); -return 0 ; + +out: +strbuf_release (type); +strbuf_release (charset); +strbuf_release (msg); + +va_start (ap, fmt); +die_message_routine (fmt, ap); +va_end (ap); +exit (128 ); } static int get_protocol_http_header(enum protocol_version version,
@@ -518,21 +529,21 @@ static struct discovery *discover_refs(const char *service, int for_push)case HTTP_OK :break ;case HTTP_MISSING_TARGET : -show_http_message (&type, &charset, &buffer); -die (_ ("repository '%s' not found" ), -transport_anonymize_url (url.buf)); +show_http_message_fatal (&type, &charset, &buffer, +_ ("repository '%s' not found" ), +transport_anonymize_url (url.buf));case HTTP_NOAUTH : -show_http_message (&type, &charset, &buffer); -die (_ ("Authentication failed for '%s'" ), -transport_anonymize_url (url.buf)); +show_http_message_fatal (&type, &charset, &buffer, +_ ("Authentication failed for '%s'" ), +transport_anonymize_url (url.buf));case HTTP_NOMATCHPUBLICKEY : -show_http_message (&type, &charset, &buffer); -die (_ ("unable to access '%s' with http.pinnedPubkey configuration: %s" ), -transport_anonymize_url (url.buf), curl_errorstr); +show_http_message_fatal (&type, &charset, &buffer, +_ ("unable to access '%s' with http.pinnedPubkey configuration: %s" ), +transport_anonymize_url (url.buf), curl_errorstr);default : -show_http_message (&type, &charset, &buffer); -die (_ ("unable to access '%s': %s" ), -transport_anonymize_url (url.buf), curl_errorstr); +show_http_message_fatal (&type, &charset, &buffer, +_ ("unable to access '%s': %s" ), +transport_anonymize_url (url.buf), curl_errorstr); }if (options.verbosity && !starts_with (refs_url.buf, url.buf)) {
Signature · 1 line
gitgitgadget
PATCH v3 3/3 http: add support for HTTP 429 rate limit retries
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
Add retry logic for HTTP 429 (Too Many Requests) responses to handle server-side rate limiting gracefully. When Git's HTTP client receives a 429 response, it can now automatically retry the request after an appropriate delay, respecting the server's rate limits.
The implementation supports the RFC-compliant Retry-After header in both delay-seconds (integer) and HTTP-date (RFC 2822) formats. If a past date is provided, Git retries immediately without waiting.
Retry behavior is controlled by three new configuration options (http.maxRetries, http.retryAfter, and http.maxRetryTime) which are documented in git-config(1).
The retry logic implements a fail-fast approach: if any delay (whether from server header or configuration) exceeds maxRetryTime, Git fails immediately with a clear error message rather than capping the delay. This provides better visibility into rate limiting issues.
The implementation includes extensive test coverage for basic retry behavior, Retry-After header formats (integer and HTTP-date), configuration combinations, maxRetryTime limits, invalid header handling, environment variable overrides, and edge cases.
--- Documentation/config/http.adoc | 23 +++ git-curl-compat.h | 8 + http.c | 190 +++++++++++++++++++++-- http.h | 2 + remote-curl.c | 4 + t/lib-httpd.sh | 1 + t/lib-httpd/apache.conf | 8 + t/lib-httpd/http-429.sh | 98 ++++++++++++ t/meson.build | 1 + t/t5584-http-429-retry.sh | 266 +++++++++++++++++++++++++++++++++ 10 files changed, 589 insertions(+), 12 deletions(-) create mode 100644 t/lib-httpd/http-429.sh create mode 100755 t/t5584-http-429-retry.sh
Documentation/config/http.adoc +23 -0
@@ -315,6 +315,29 @@ http.keepAliveCount:: unset, curl's default value is used. Can be overridden by the `GIT_HTTP_KEEPALIVE_COUNT` environment variable. +http.retryAfter:: + Default wait time in seconds before retrying when a server returns + HTTP 429 (Too Many Requests) without a Retry-After header. + Defaults to 0 (retry immediately). When a Retry-After header is + present, its value takes precedence over this setting. Can be + overridden by the `GIT_HTTP_RETRY_AFTER` environment variable. + See also `http.maxRetries` and `http.maxRetryTime`. + +http.maxRetries:: + Maximum number of times to retry after receiving HTTP 429 (Too Many + Requests) responses. Set to 0 (the default) to disable retries. + Can be overridden by the `GIT_HTTP_MAX_RETRIES` environment variable. + See also `http.retryAfter` and `http.maxRetryTime`. + +http.maxRetryTime:: + Maximum time in seconds to wait for a single retry attempt when + handling HTTP 429 (Too Many Requests) responses. If the server + requests a delay (via Retry-After header) or if `http.retryAfter` + is configured with a value that exceeds this maximum, Git will fail + immediately rather than waiting. Default is 300 seconds (5 minutes). + Can be overridden by the `GIT_HTTP_MAX_RETRY_TIME` environment + variable. See also `http.retryAfter` and `http.maxRetries`. + http.noEPSV:: A boolean which disables using of EPSV ftp command by curl. This can be helpful with some "poor" ftp servers which don't
git-curl-compat.h +8 -0
@@ -37,6 +37,14 @@#define GIT_CURL_NEED_TRANSFER_ENCODING_HEADER #endif +/** +* CURLINFO_RETRY_AFTER was added in 7.66.0, released in September 2019. +* It allows curl to automatically parse Retry-After headers. +*/ +#if LIBCURL_VERSION_NUM >=0x074200 +#define GIT_CURL_HAVE_CURLINFO_RETRY_AFTER 1 +#endif +/** * CURLOPT_PROTOCOLS_STR and CURLOPT_REDIR_PROTOCOLS_STR were added in 7.85.0, * released in August 2022.
http.c +178 -12
305 lines not on this page · raw message
http.h +2 -0
@@ -20,6 +20,7 @@ struct slot_results {long http_code;long auth_avail;long http_connectcode; +long retry_after; };struct active_request_slot {
@@ -167,6 +168,7 @@ struct http_get_options {#define HTTP_REAUTH 4#define HTTP_NOAUTH 5#define HTTP_NOMATCHPUBLICKEY 6 +#define HTTP_RATE_LIMITED 7/* * Requests a URL and stores the result in a strbuf.
remote-curl.c +4 -0
@@ -540,6 +540,10 @@ static struct discovery *discover_refs(const char *service, int for_push)show_http_message_fatal (&type, &charset, &buffer,_ ("unable to access '%s' with http.pinnedPubkey configuration: %s" ),transport_anonymize_url (url.buf), curl_errorstr); +case HTTP_RATE_LIMITED : +show_http_message_fatal (&type, &charset, &buffer, +_ ("rate limited by '%s', please try again later" ), +transport_anonymize_url (url.buf));default : show_http_message_fatal(&type, &charset, &buffer, _("unable to access '%s': %s"),
t/lib-httpd.sh +1 -0
@@ -167,6 +167,7 @@ prepare_httpd() { install_script error.sh install_script apply-one-time-script.sh install_script nph-custom-auth.sh + install_script http-429.sh ln -s "$LIB_HTTPD_MODULE_PATH" "$HTTPD_ROOT_PATH/modules"
t/lib-httpd/apache.conf +8 -0
@@ -139,6 +139,10 @@ SetEnv PERL_PATH ${PERL_PATH} SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} SetEnv GIT_HTTP_EXPORT_ALL </LocationMatch> +<LocationMatch /http_429/> + SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} + SetEnv GIT_HTTP_EXPORT_ALL +</LocationMatch> <LocationMatch /smart_v0/> SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} SetEnv GIT_HTTP_EXPORT_ALL
@@ -160,6 +164,7 @@ ScriptAlias /broken_smart/ broken-smart-http.sh/ ScriptAlias /error_smart/ error-smart-http.sh/ ScriptAlias /error/ error.sh/ ScriptAliasMatch /one_time_script/(.*) apply-one-time-script.sh/$1 +ScriptAliasMatch /http_429/(.*) http-429.sh/$1 ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1 <Directory ${GIT_EXEC_PATH}> Options FollowSymlinks
@@ -185,6 +190,9 @@ ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1 <Files apply-one-time-script.sh> Options ExecCGI </Files> +<Files http-429.sh> + Options ExecCGI +</Files> <Files ${GIT_EXEC_PATH}/git-http-backend> Options ExecCGI </Files>
t/lib-httpd/http-429.sh new file +98 -0
98 lines not on this page · raw message
t/meson.build +1 -0
@@ -703,6 +703,7 @@ integration_tests = [ 't5581-http-curl-verbose.sh', 't5582-fetch-negative-refspec.sh', 't5583-push-branches.sh', + 't5584-http-429-retry.sh', 't5600-clone-fail-cleanup.sh', 't5601-clone.sh', 't5602-clone-remote-exec.sh',
t/t5584-http-429-retry.sh new file +266 -0
266 lines not on this page · raw message
Signature · 1 line
gitgitgadget
Re: [PATCH v3 1/3] strbuf: fix incorrect alloc size in strbuf_reencode()
In reply to Vaidas Pilkauskas via GitGitGadget"Vaidas Pilkauskas via GitGitGadget" <[email protected]>
Quoted 12 lines · writes:
From: Vaidas Pilkauskas <[email protected]>
The
strbuf_reencode()function incorrectly passes the string length as the allocation size tostrbuf_attach(), when it should pass length + 1 to account for the null terminator.The
reencode_string_len()function allocates len + 1 bytes (including the null terminator) and returns the string length (excluding the null terminator) via the len parameter. However,strbuf_reencode()then callsstrbuf_attach()with this length value as both the len and alloc parameters:
Quoted 6 lines
strbuf_attach(sb, out, len, len);This is incorrect because strbuf_attach()'s alloc parameter should reflect the actual allocated buffer size, which includes space for the null terminator. This could lead to incorrect memory management in code that relies on
sb->allocbeing accurate.
I do agree that setting the correct number to .alloc member is a good thing to do, but I am afraid that the above characterization of a potential problem is incorrect.
If we were to extend the resulting strbuf further (by e.g., appending to it), we might end up reallocating the buffer a bit prematurely by one byte before it actually fills up, but the reallocation would be done by giving the piece of memory pointed at by "out" here to realloc(3), so the wrong value of "alloc" would not lead to incorrect memory management at all.
Upon further inspection, we see something else interesting. The strbuf_attach() function, immediately after initializing sb with the new values of buf/len/alloc, calls strbuf_grow(sb, 0) and triggers the ALLOC_GROW() growth thanks to this under specification. By the time the control returns to the caller, the sb->alloc would be (((len)+16)*3/2), not (len+1), and it records the actual allocation size. So there is no "could lead to incorrect memory management" at all, but this incorrect number forces us to always reallocate immediately after the strbuf_attach() call, which is a waste when we are not going to further extend the strbuf returned by this function.
And that is a very good reason to make this fix worth doing.
Quoted 3 lines
Fix by passing len + 1 as the alloc parameter:
strbuf_attach(sb, out, len, len + 1);
I wonder how widespread this off-by-one error is. Shouldn't strbuf_attach() be doing some sanity checking of its parameters?
void strbuf_attach(struct strbuf *sb, void *buf, size_t len, size_t alloc)
{
strbuf_release(sb);
sb->buf = buf;
sb->len = len;
sb->alloc = alloc;
strbuf_grow(sb, 0);
sb->buf[sb->len] = '\0';
}Given the above code, it is clear that alloc must be at least as big as (len + 1), and the strbuf_grow(sb, 0) in between is papering over problems (at least it is doing so here for the caller you corrected).
Perhaps we want to replace the call to strbuf_grow(sb, 0) with something like
if (alloc <= len)
BUG("alloc must be larger than len");instead? The log message of 917c9a71 (New strbuf APIs: splice and attach., 2007-09-15) is worth reading, but it is an iffy logic that depends too much (at least for my taste) on what strbuf_grow(sb, 0) actually does ;-).
Quoted 17 lines
Signed-off-by Vaidas Pilkauskas <[email protected]>--- strbuf.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)strbuf.c +1
-1@@ -168,7 +168,7 @@ int strbuf_reencode(struct strbuf *sb, const char *from, const char *to)if (!out)return -1 ; -strbuf_attach (sb, out, len, len); +strbuf_attach (sb, out, len, len +1 );return 0 ; }
Re: [PATCH v3 1/3] strbuf: fix incorrect alloc size in strbuf_reencode()
Quoted 41 lines · Junio C Hamano
Quoted 12 lines
From: Vaidas Pilkauskas <[email protected]>
The
strbuf_reencode()function incorrectly passes the string length as the allocation size tostrbuf_attach(), when it should pass length + 1 to account for the null terminator.The
reencode_string_len()function allocates len + 1 bytes (including the null terminator) and returns the string length (excluding the null terminator) via the len parameter. However,strbuf_reencode()then callsstrbuf_attach()with this length value as both the len and alloc parameters:Quoted 6 lines
strbuf_attach(sb, out, len, len);This is incorrect because strbuf_attach()'s alloc parameter should reflect the actual allocated buffer size, which includes space for the null terminator. This could lead to incorrect memory management in code that relies on
sb->allocbeing accurate.I do agree that setting the correct number to .alloc member is a good thing to do, but I am afraid that the above characterization of a potential problem is incorrect.
If we were to extend the resulting strbuf further (by e.g., appending to it), we might end up reallocating the buffer a bit prematurely by one byte before it actually fills up, but the reallocation would be done by giving the piece of memory pointed at by "out" here to realloc(3), so the wrong value of "alloc" would not lead to incorrect memory management at all.
Upon further inspection, we see something else interesting. The
strbuf_attach()function, immediately after initializing sb with the new values of buf/len/alloc, callsstrbuf_grow(sb, 0)and triggers theALLOC_GROW()growth thanks to this under specification. By the time the control returns to the caller, thesb->allocwould be (((len)+16)*3/2), not (len+1), and it records the actual allocation size. So there is no "could lead to incorrect memory management" at all, but this incorrect number forces us to always reallocate immediately after thestrbuf_attach()call, which is a waste when we are not going to further extend the strbuf returned by this function.
Quoted 1 line
And that is a very good reason to make this fix worth doing.
I agree that this is incorrect characterization. What about something like this:
strbuf: pass correct alloc to strbuf_attach() in strbuf_reencode()
reencode_string_len() allocates len+1 bytes (including the NUL) and returns the string length in len. strbuf_reencode() was calling strbuf_attach(sb, out, len, len), so alloc was one byte too small.
strbuf_attach() then calls strbuf_grow(sb, 0). With alloc < len+1, ALLOC_GROW always reallocates, so we reallocated immediately after attach even when the strbuf was not extended further. Pass len+1 as the alloc argument so the existing buffer is reused and the reallocation is avoided.
Quoted 25 lines
Quoted 3 lines
Fix by passing len + 1 as the alloc parameter:
strbuf_attach(sb, out, len, len + 1);I wonder how widespread this off-by-one error is. Shouldn't
strbuf_attach()be doing some sanity checking of its parameters?void strbuf_attach(struct strbuf *sb, void *buf, size_t len, size_t alloc) { strbuf_release(sb); sb->buf = buf; sb->len = len; sb->alloc = alloc; strbuf_grow(sb, 0); sb->buf[sb->len] = '\0'; }Given the above code, it is clear that alloc must be at least as big as (len + 1), and the
strbuf_grow(sb, 0)in between is papering over problems (at least it is doing so here for the caller you corrected).Perhaps we want to replace the call to
strbuf_grow(sb, 0)with something like
Quoted 7 lines
if (alloc <= len)
BUG("alloc must be larger than len");instead? The log message of
917c9a71(New strbuf APIs: splice and attach., 2007-09-15) is worth reading, but it is an iffy logic that depends too much (at least for my taste) on whatstrbuf_grow(sb, 0)actually does ;-).
I'll send patches, one to clean up call sites (there aren't too many - 7 places) and another to add BUG() check to enforce the contract.
Thanks, Junio, for the review!
Quoted 18 lines
Quoted 18 lines
Signed-off-by Vaidas Pilkauskas <[email protected]>--- strbuf.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)strbuf.c +1
-1@@ -168,7 +168,7 @@ int strbuf_reencode(struct strbuf *sb, const char *from, const char *to)if (!out)return -1 ; -strbuf_attach (sb, out, len, len); +strbuf_attach (sb, out, len, len +1 );return 0 ; }
v4 cover http: add support for HTTP 429 rate limit retries
In reply to Vaidas Pilkauskas via GitGitGadgetChanges since v3:
* Clean up of all strbuf_attach() call sites
* Add strbuf_attach() contract enforcement via BUG()
Changes since v2:
* New preparatory patch: Introduced show_http_message_fatal() helper function to reduce code duplication in remote-curl.c (suggested by Taylor Blau)
* Removed specific HTTP_RATE_LIMITED error handling from http-push.c and http-walker.c for the obsolete "dumb" protocol, allowing generic error handling to take over (suggested by Jeff King)
* Added support for CURLINFO_RETRY_AFTER on curl >= 7.66.0, falling back to manual header parsing on older versions
* Simplified retry/delay architecture: replaced complex non-blocking "delayed slot" mechanism with simple blocking sleep() call in the retry loop, removing ~66 lines of timing logic (suggested by Jeff King)
* Fixed Retry-After: 0 handling to allow immediate retry as specified by RFC 9110
* Changed http.retryAfter default from -1 to 0, so Git will retry immediately when encountering HTTP 429 without a Retry-After header, rather than failing with a configuration error
* Improved error messages: shortened to be more concise
* Fixed coding style issues: removed unnecessary curly braces, changed x == 0 to !x (per CodingGuidelines)
* Improved test portability: replaced non-portable date(1) commands with test-tool date, added nanosecond-precision timing with getnanos, replaced cut(1) with POSIX shell parameter expansion
* Split out strbuf.c bugfix into separate preparatory patch (the strbuf_reencode alloc size fix is unrelated to HTTP 429 support)
* Squashed separate trace2 logging patch into main HTTP 429 retry support commit
* Kept header_is_last_match assignment for Retry-After to prevent incorrect handling of HTTP header continuation lines
The implementation includes:
1. A bug fix in strbuf_reencode() that corrects the allocation size passed to strbuf_attach(), passing len+1 instead of len so that the existing buffer is reused rather than immediately reallocated.
2. A cleanup of all strbuf_attach() call sites that were passing alloc == len, leaving no room for the NUL terminator. Sites with a known-NUL-terminated buffer now pass len+1; sites where the source buffer has no trailing NUL (ll_merge output) are converted to use strbuf_add() instead.
3. A hardening of strbuf_attach() itself: the internal strbuf_grow() that silently papered over incorrect alloc values is replaced with an explicit BUG() check, enforcing the documented contract that alloc must be greater than len.
4. A new show_http_message_fatal() helper in remote-curl.c that combines the repeated pattern of show_http_message() followed by die() into a single NORETURN function, reducing boilerplate at existing call sites and providing a clean hook for the retry logic.
5. The main feature: HTTP 429 retry logic with support for the Retry-After header (both delay-seconds and HTTP-date formats), configurable via http.maxRetries, http.retryAfter, and http.maxRetryTime options. If any computed delay exceeds maxRetryTime the request fails immediately with a clear diagnostic rather than capping and retrying silently.
Vaidas Pilkauskas (5):
strbuf: pass correct alloc to strbuf_attach() in strbuf_reencode()
strbuf_attach: fix all call sites to pass correct alloc
strbuf: replace strbuf_grow() in strbuf_attach() with BUG() check
remote-curl: introduce show_http_message_fatal() helper
http: add support for HTTP 429 rate limit retries
Documentation/config/http.adoc | 23 +++ apply.c | 3 +- builtin/am.c | 2 +- builtin/fast-import.c | 2 +- git-curl-compat.h | 8 + http.c | 190 +++++++++++++++++++++-- http.h | 2 + mailinfo.c | 2 +- refs/files-backend.c | 2 +- remote-curl.c | 49 +++--- rerere.c | 3 +- strbuf.c | 5 +- t/lib-httpd.sh | 1 + t/lib-httpd/apache.conf | 8 + t/lib-httpd/http-429.sh | 98 ++++++++++++ t/meson.build | 1 + t/t5584-http-429-retry.sh | 266 +++++++++++++++++++++++++++++++++ trailer.c | 2 +- 18 files changed, 629 insertions(+), 38 deletions(-) create mode 100644 t/lib-httpd/http-429.sh create mode 100755 t/t5584-http-429-retry.sh
base-commit: 73fd77805fc6406f31c36212846d9e2541d19321
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-2008%2Fvaidas-shopify%2Fretry-after-v4
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-2008/vaidas-shopify/retry-after-v4
Pull-Request: https://github.com/gitgitgadget/git/pull/2008
Range-diff vs v3:
1: 821043c664 ! 1: a3386f5b56 strbuf: fix incorrect alloc size in strbuf_reencode()
@@ Metadata
Author: Vaidas Pilkauskas <[email protected]>
## Commit message ##
- strbuf: fix incorrect alloc size in strbuf_reencode()
+ strbuf: pass correct alloc to strbuf_attach() in strbuf_reencode()
- The strbuf_reencode() function incorrectly passes the string length
- as the allocation size to strbuf_attach(), when it should pass
- length + 1 to account for the null terminator.
+ reencode_string_len() allocates len+1 bytes (including the NUL) and
+ returns the string length in len. strbuf_reencode() was calling
+ strbuf_attach(sb, out, len, len), so alloc was one byte too small.
- The reencode_string_len() function allocates len + 1 bytes (including
- the null terminator) and returns the string length (excluding the null
- terminator) via the len parameter. However, strbuf_reencode() then
- calls strbuf_attach() with this length value as both the len and alloc
- parameters:
-
- strbuf_attach(sb, out, len, len);
-
- This is incorrect because strbuf_attach()'s alloc parameter should
- reflect the actual allocated buffer size, which includes space for the
- null terminator. This could lead to incorrect memory management in code
- that relies on sb->alloc being accurate.
-
- Fix by passing len + 1 as the alloc parameter:
-
- strbuf_attach(sb, out, len, len + 1);
+ strbuf_attach() then calls strbuf_grow(sb, 0). With alloc < len+1,
+ ALLOC_GROW always reallocates, so we reallocated immediately after
+ attach even when the strbuf was not extended further. Pass len+1 as
+ the alloc argument so the existing buffer is reused and the
+ reallocation is avoided.
-: ---------- > 2: f48b1f07c4 strbuf_attach: fix all call sites to pass correct alloc -: ---------- > 3: 557fd77444 strbuf: replace strbuf_grow() in strbuf_attach() with BUG() check 2: 3653067f0e = 4: 3a39dc9e39 remote-curl: introduce show_http_message_fatal() helper 3: 3cece62a63 = 5: 5e0f4a56ef http: add support for HTTP 429 rate limit retries
Signature · 1 line
gitgitgadget
PATCH v4 1/5 strbuf: pass correct alloc to strbuf_attach() in strbuf_reencode()
From: Vaidas Pilkauskas <[email protected]>
reencode_string_len() allocates len+1 bytes (including the NUL) and returns the string length in len. strbuf_reencode() was calling strbuf_attach(sb, out, len, len), so alloc was one byte too small.
strbuf_attach() then calls strbuf_grow(sb, 0). With alloc < len+1, ALLOC_GROW always reallocates, so we reallocated immediately after attach even when the strbuf was not extended further. Pass len+1 as the alloc argument so the existing buffer is reused and the reallocation is avoided.
--- strbuf.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
strbuf.c +1 -1
@@ -168,7 +168,7 @@ int strbuf_reencode(struct strbuf *sb, const char *from, const char *to)if (!out)return -1 ; -strbuf_attach (sb, out, len, len); +strbuf_attach (sb, out, len, len +1 );return 0 ; }
Signature · 1 line
gitgitgadget
PATCH v4 2/5 strbuf_attach: fix all call sites to pass correct alloc
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
strbuf_attach(sb, buf, len, alloc) requires alloc > len (the buffer must have at least len+1 bytes to hold the NUL). Several call sites passed alloc == len, relying on strbuf_grow(sb, 0) inside strbuf_attach to reallocate. Prepare for changing that by fixing call sites to pass the correct alloc.
- mailinfo, am, refs/files-backend, fast-import, trailer: pass len+1 when the buffer is a NUL-terminated string (or from strbuf_detach).
- rerere, apply: ll_merge returns a buffer with exactly result.size bytes (no extra NUL). Use strbuf_add() to copy and NUL-terminate into the strbuf, then free the merge result, so alloc is correct.
--- apply.c | 3 ++- builtin/am.c | 2 +- builtin/fast-import.c | 2 +- mailinfo.c | 2 +- refs/files-backend.c | 2 +- rerere.c | 3 ++- trailer.c | 2 +- 7 files changed, 9 insertions(+), 7 deletions(-)
apply.c +2 -1
@@ -3589,7 +3589,8 @@ static int three_way_merge(struct apply_state *state,return -1 ; }image_clear (image); -strbuf_attach (&image->buf, result.ptr, result.size, result.size); +strbuf_add (&image->buf, result.ptr, result.size); +free (result.ptr);return status; }
builtin/am.c +1 -1
@@ -1188,7 +1188,7 @@ static void am_append_signoff(struct am_state *state) {struct strbuf sb =STRBUF_INIT ; -strbuf_attach (&sb, state->msg, state->msg_len, state->msg_len); +strbuf_attach (&sb, state->msg, state->msg_len, state->msg_len +1 );append_signoff (&sb,0 ,0 ); state->msg =strbuf_detach (&sb, &state->msg_len); }
builtin/fast-import.c +1 -1
@@ -3246,7 +3246,7 @@ static void cat_blob(struct object_entry *oe, struct object_id *oid)cat_blob_write ("\n" ,1 );if (oe && oe->pack_id == pack_id) { last_blob.offset = oe->idx.offset; -strbuf_attach (&last_blob.data, buf, size, size); +strbuf_attach (&last_blob.data, buf, size, size +1 ); last_blob.depth = oe->depth; }else free (buf);
mailinfo.c +1 -1
@@ -470,7 +470,7 @@ static int convert_to_utf8(struct mailinfo *mi,return error ("cannot convert from %s to %s" , charset, mi->metainfo_charset); } -strbuf_attach (line, out, out_len, out_len); +strbuf_attach (line, out, out_len, out_len +1 );return 0 ; }
refs/files-backend.c +1 -1
@@ -1806,7 +1806,7 @@ static int commit_ref(struct ref_lock *lock)size_t len =strlen (path);struct strbuf sb_path =STRBUF_INIT ; -strbuf_attach (&sb_path, path, len, len); +strbuf_attach (&sb_path, path, len, len +1 );/* * If this fails, commit_lock_file() will also fail
rerere.c +2 -1
@@ -1031,7 +1031,8 @@ static int handle_cache(struct index_state *istate,else io.io.output =NULL ;strbuf_init (&io.input,0 ); -strbuf_attach (&io.input, result.ptr, result.size, result.size); +strbuf_add (&io.input, result.ptr, result.size); +free (result.ptr);/* * Grab the conflict ID and optionally write the original
trailer.c +1 -1
@@ -1009,7 +1009,7 @@ static struct trailer_block *trailer_block_get(const struct process_trailer_optifor (ptr = trailer_lines; *ptr; ptr++) {if (last &&isspace ((*ptr)->buf[0 ])) {struct strbuf sb =STRBUF_INIT ; -strbuf_attach (&sb, *last,strlen (*last),strlen (*last)); +strbuf_attach (&sb, *last,strlen (*last),strlen (*last) +1 );strbuf_addbuf (&sb, *ptr); *last =strbuf_detach (&sb,NULL );continue ;
Signature · 1 line
gitgitgadget
PATCH v4 3/5 strbuf: replace strbuf_grow() in strbuf_attach() with BUG() check
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
strbuf_attach() documents that alloc must be larger than len, as the buffer must have room for the NUL terminator. Replace the strbuf_grow(sb, 0) call, which was silently reallocating when alloc <= len, with an explicit BUG() to enforce this contract and write the NUL terminator directly.
--- strbuf.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
strbuf.c +2 -1
@@ -95,11 +95,12 @@ char *strbuf_detach(struct strbuf *sb, size_t *sz)void strbuf_attach (struct strbuf *sb,void *buf,size_t len,size_t alloc) { +if (alloc <= len) +BUG ("alloc must be larger than len" );strbuf_release (sb); sb->buf = buf; sb->len = len; sb->alloc = alloc; -strbuf_grow (sb,0 ); sb->buf[sb->len] ='\0' ; }
Signature · 1 line
gitgitgadget
PATCH v4 4/5 remote-curl: introduce show_http_message_fatal() helper
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
Several code paths in remote-curl.c follow the same pattern of calling show_http_message() to display server error messages followed by die() to terminate with an error. This duplication makes the code more verbose and harder to maintain.
Introduce a new show_http_message_fatal() helper function that combines these two operations. This function:
1. Displays any HTTP error message from the server via show_http_message()
2. Calls die() with the provided error message
3. Returns NORETURN to help the compiler with control flow analysis
Refactor existing call sites in remote-curl.c to use this new helper, reducing code duplication and improving readability. This pattern will also be used by upcoming HTTP 429 rate limiting support.
--- remote-curl.c | 45 ++++++++++++++++++++++++++++----------------- 1 file changed, 28 insertions(+), 17 deletions(-)
remote-curl.c +28 -17
@@ -367,23 +367,25 @@ static void free_discovery(struct discovery *d) } } -static int show_http_message (struct strbuf *type,struct strbuf *charset, -struct strbuf *msg) +static NORETURN void show_http_message_fatal(struct strbuf *type, struct strbuf *charset, +struct strbuf *msg,const char *fmt, ...) {const char *p, *eol; +va_list ap; +report_fn die_message_routine =get_die_message_routine ();/* * We only show text/plain parts, as other types are likely * to be ugly to look at on the user's terminal. */ if (strcmp (type->buf,"text/plain" )) -return -1 ; + goto out;if (charset->len)strbuf_reencode (msg, charset->buf,get_log_output_encoding ());strbuf_trim (msg);if (!msg->len) -return -1 ; + goto out; p = msg->buf; do {
@@ -391,7 +393,16 @@ static int show_http_message(struct strbuf *type, struct strbuf *charset,fprintf (stderr,"remote: %.*s\n" , (int )(eol - p), p); p = eol +1 ; }while (*eol); -return 0 ; + +out: +strbuf_release (type); +strbuf_release (charset); +strbuf_release (msg); + +va_start (ap, fmt); +die_message_routine (fmt, ap); +va_end (ap); +exit (128 ); } static int get_protocol_http_header(enum protocol_version version,
@@ -518,21 +529,21 @@ static struct discovery *discover_refs(const char *service, int for_push)case HTTP_OK :break ;case HTTP_MISSING_TARGET : -show_http_message (&type, &charset, &buffer); -die (_ ("repository '%s' not found" ), -transport_anonymize_url (url.buf)); +show_http_message_fatal (&type, &charset, &buffer, +_ ("repository '%s' not found" ), +transport_anonymize_url (url.buf));case HTTP_NOAUTH : -show_http_message (&type, &charset, &buffer); -die (_ ("Authentication failed for '%s'" ), -transport_anonymize_url (url.buf)); +show_http_message_fatal (&type, &charset, &buffer, +_ ("Authentication failed for '%s'" ), +transport_anonymize_url (url.buf));case HTTP_NOMATCHPUBLICKEY : -show_http_message (&type, &charset, &buffer); -die (_ ("unable to access '%s' with http.pinnedPubkey configuration: %s" ), -transport_anonymize_url (url.buf), curl_errorstr); +show_http_message_fatal (&type, &charset, &buffer, +_ ("unable to access '%s' with http.pinnedPubkey configuration: %s" ), +transport_anonymize_url (url.buf), curl_errorstr);default : -show_http_message (&type, &charset, &buffer); -die (_ ("unable to access '%s': %s" ), -transport_anonymize_url (url.buf), curl_errorstr); +show_http_message_fatal (&type, &charset, &buffer, +_ ("unable to access '%s': %s" ), +transport_anonymize_url (url.buf), curl_errorstr); }if (options.verbosity && !starts_with (refs_url.buf, url.buf)) {
Signature · 1 line
gitgitgadget
PATCH v4 5/5 http: add support for HTTP 429 rate limit retries
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
Add retry logic for HTTP 429 (Too Many Requests) responses to handle server-side rate limiting gracefully. When Git's HTTP client receives a 429 response, it can now automatically retry the request after an appropriate delay, respecting the server's rate limits.
The implementation supports the RFC-compliant Retry-After header in both delay-seconds (integer) and HTTP-date (RFC 2822) formats. If a past date is provided, Git retries immediately without waiting.
Retry behavior is controlled by three new configuration options (http.maxRetries, http.retryAfter, and http.maxRetryTime) which are documented in git-config(1).
The retry logic implements a fail-fast approach: if any delay (whether from server header or configuration) exceeds maxRetryTime, Git fails immediately with a clear error message rather than capping the delay. This provides better visibility into rate limiting issues.
The implementation includes extensive test coverage for basic retry behavior, Retry-After header formats (integer and HTTP-date), configuration combinations, maxRetryTime limits, invalid header handling, environment variable overrides, and edge cases.
--- Documentation/config/http.adoc | 23 +++ git-curl-compat.h | 8 + http.c | 190 +++++++++++++++++++++-- http.h | 2 + remote-curl.c | 4 + t/lib-httpd.sh | 1 + t/lib-httpd/apache.conf | 8 + t/lib-httpd/http-429.sh | 98 ++++++++++++ t/meson.build | 1 + t/t5584-http-429-retry.sh | 266 +++++++++++++++++++++++++++++++++ 10 files changed, 589 insertions(+), 12 deletions(-) create mode 100644 t/lib-httpd/http-429.sh create mode 100755 t/t5584-http-429-retry.sh
Documentation/config/http.adoc +23 -0
@@ -315,6 +315,29 @@ http.keepAliveCount:: unset, curl's default value is used. Can be overridden by the `GIT_HTTP_KEEPALIVE_COUNT` environment variable. +http.retryAfter:: + Default wait time in seconds before retrying when a server returns + HTTP 429 (Too Many Requests) without a Retry-After header. + Defaults to 0 (retry immediately). When a Retry-After header is + present, its value takes precedence over this setting. Can be + overridden by the `GIT_HTTP_RETRY_AFTER` environment variable. + See also `http.maxRetries` and `http.maxRetryTime`. + +http.maxRetries:: + Maximum number of times to retry after receiving HTTP 429 (Too Many + Requests) responses. Set to 0 (the default) to disable retries. + Can be overridden by the `GIT_HTTP_MAX_RETRIES` environment variable. + See also `http.retryAfter` and `http.maxRetryTime`. + +http.maxRetryTime:: + Maximum time in seconds to wait for a single retry attempt when + handling HTTP 429 (Too Many Requests) responses. If the server + requests a delay (via Retry-After header) or if `http.retryAfter` + is configured with a value that exceeds this maximum, Git will fail + immediately rather than waiting. Default is 300 seconds (5 minutes). + Can be overridden by the `GIT_HTTP_MAX_RETRY_TIME` environment + variable. See also `http.retryAfter` and `http.maxRetries`. + http.noEPSV:: A boolean which disables using of EPSV ftp command by curl. This can be helpful with some "poor" ftp servers which don't
git-curl-compat.h +8 -0
@@ -37,6 +37,14 @@#define GIT_CURL_NEED_TRANSFER_ENCODING_HEADER #endif +/** +* CURLINFO_RETRY_AFTER was added in 7.66.0, released in September 2019. +* It allows curl to automatically parse Retry-After headers. +*/ +#if LIBCURL_VERSION_NUM >=0x074200 +#define GIT_CURL_HAVE_CURLINFO_RETRY_AFTER 1 +#endif +/** * CURLOPT_PROTOCOLS_STR and CURLOPT_REDIR_PROTOCOLS_STR were added in 7.85.0, * released in August 2022.
http.c +178 -12
305 lines not on this page · raw message
http.h +2 -0
@@ -20,6 +20,7 @@ struct slot_results {long http_code;long auth_avail;long http_connectcode; +long retry_after; };struct active_request_slot {
@@ -167,6 +168,7 @@ struct http_get_options {#define HTTP_REAUTH 4#define HTTP_NOAUTH 5#define HTTP_NOMATCHPUBLICKEY 6 +#define HTTP_RATE_LIMITED 7/* * Requests a URL and stores the result in a strbuf.
remote-curl.c +4 -0
@@ -540,6 +540,10 @@ static struct discovery *discover_refs(const char *service, int for_push)show_http_message_fatal (&type, &charset, &buffer,_ ("unable to access '%s' with http.pinnedPubkey configuration: %s" ),transport_anonymize_url (url.buf), curl_errorstr); +case HTTP_RATE_LIMITED : +show_http_message_fatal (&type, &charset, &buffer, +_ ("rate limited by '%s', please try again later" ), +transport_anonymize_url (url.buf));default : show_http_message_fatal(&type, &charset, &buffer, _("unable to access '%s': %s"),
t/lib-httpd.sh +1 -0
@@ -167,6 +167,7 @@ prepare_httpd() { install_script error.sh install_script apply-one-time-script.sh install_script nph-custom-auth.sh + install_script http-429.sh ln -s "$LIB_HTTPD_MODULE_PATH" "$HTTPD_ROOT_PATH/modules"
t/lib-httpd/apache.conf +8 -0
@@ -139,6 +139,10 @@ SetEnv PERL_PATH ${PERL_PATH} SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} SetEnv GIT_HTTP_EXPORT_ALL </LocationMatch> +<LocationMatch /http_429/> + SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} + SetEnv GIT_HTTP_EXPORT_ALL +</LocationMatch> <LocationMatch /smart_v0/> SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} SetEnv GIT_HTTP_EXPORT_ALL
@@ -160,6 +164,7 @@ ScriptAlias /broken_smart/ broken-smart-http.sh/ ScriptAlias /error_smart/ error-smart-http.sh/ ScriptAlias /error/ error.sh/ ScriptAliasMatch /one_time_script/(.*) apply-one-time-script.sh/$1 +ScriptAliasMatch /http_429/(.*) http-429.sh/$1 ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1 <Directory ${GIT_EXEC_PATH}> Options FollowSymlinks
@@ -185,6 +190,9 @@ ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1 <Files apply-one-time-script.sh> Options ExecCGI </Files> +<Files http-429.sh> + Options ExecCGI +</Files> <Files ${GIT_EXEC_PATH}/git-http-backend> Options ExecCGI </Files>
t/lib-httpd/http-429.sh new file +98 -0
98 lines not on this page · raw message
t/meson.build +1 -0
@@ -704,6 +704,7 @@ integration_tests = [ 't5581-http-curl-verbose.sh', 't5582-fetch-negative-refspec.sh', 't5583-push-branches.sh', + 't5584-http-429-retry.sh', 't5600-clone-fail-cleanup.sh', 't5601-clone.sh', 't5602-clone-remote-exec.sh',
t/t5584-http-429-retry.sh new file +266 -0
266 lines not on this page · raw message
Signature · 1 line
gitgitgadget
Re: [PATCH v4 2/5] strbuf_attach: fix all call sites to pass correct alloc
In reply to Vaidas Pilkauskas via GitGitGadget"Vaidas Pilkauskas via GitGitGadget" <[email protected]>
Quoted 2 lines · writes:
- mailinfo, am, refs/files-backend, fast-import, trailer: pass len+1 when the buffer is a NUL-terminated string (or from
strbuf_detach).
These are good.
Quoted 3 lines
- rerere, apply:
ll_mergereturns a buffer with exactly result.size bytes (no extra NUL). Usestrbuf_add()to copy and NUL-terminate into the strbuf, then free the merge result, so alloc is correct.
I am not sure about this, because this will result in unnecessary reallocation.
For example
apply.c static int three_way_merge(struct apply_state *state, · line 3592
Quoted 1 line
apply.c +0
-1-strbuf_attach (&image->buf, result.ptr, result.size, result.size);
This would have resulted in realloc(result.ptr, result.size + X) to preserve the strbuf invariants that len + 1 <= alloc inside the strbuf_attach().
It depends on what the system allocator does, but when X is a small number, often no new memory needs to be carved out when this realloc() happens, and all that needs to happen is that the size of the memory region recorded by the system allocator is adjusted, and the program will keep using the same memory region plus X bytes out of the slop that has already been there when result.ptr was allocated. We will call realloc(), and it may result in a true allocation and copy when X is larger than the existing slop, but it may end up to be a cheap operation.
But if we rewrite it to do this ...
apply.c static int three_way_merge(struct apply_state *state, · line 3593
Quoted 2 lines
apply.c +2
-0+strbuf_add (&image->buf, result.ptr, result.size); +free (result.ptr);
... we will allocate as much as result.size and copy the bytes. Guaranteed, regardless of how much slop the system allocator left after result.ptr+result.size when it allocated result.ptr.
Of course, we could rewrite the original to
result.ptr = realloc(result.ptr, result.size + 1);
strbuf_attach(&image->buf, result.ptr, result.size, result.size + 1);which would avoid the extra allocation and copy when there is even a single byte of slop after result.ptr+result.size, but at that point, for the sake of simplicity, we may be better off with the original implementation of strbuf_attach() that automatically does that for us.
So, I dunno.
Thanks.
Re: [PATCH v4 2/5] strbuf_attach: fix all call sites to pass correct alloc
Quoted 36 lines · Junio C Hamano
Quoted 3 lines
- rerere, apply:
ll_mergereturns a buffer with exactly result.size bytes (no extra NUL). Usestrbuf_add()to copy and NUL-terminate into the strbuf, then free the merge result, so alloc is correct.I am not sure about this, because this will result in unnecessary reallocation.
For example
Quoted 1 line
-
strbuf_attach(&image->buf, result.ptr, result.size, result.size);This would have resulted in
realloc(result.ptr, result.size + X)to preserve the strbuf invariants that len + 1 <= alloc inside thestrbuf_attach().It depends on what the system allocator does, but when X is a small number, often no new memory needs to be carved out when this
realloc()happens, and all that needs to happen is that the size of the memory region recorded by the system allocator is adjusted, and the program will keep using the same memory region plus X bytes out of the slop that has already been there when result.ptr was allocated. We will callrealloc(), and it may result in a true allocation and copy when X is larger than the existing slop, but it may end up to be a cheap operation.But if we rewrite it to do this ...
apply.c static int three_way_merge(struct apply_state *state, · line 3593
Quoted 2 lines
apply.c +2
-0+strbuf_add (&image->buf, result.ptr, result.size); +free (result.ptr);... we will allocate as much as result.size and copy the bytes. Guaranteed, regardless of how much slop the system allocator left after result.ptr+result.size when it allocated result.ptr.
Of course, we could rewrite the original to
Quoted 8 lines
result.ptr = realloc(result.ptr, result.size + 1); strbuf_attach(&image->buf, result.ptr, result.size, result.size + 1);which would avoid the extra allocation and copy when there is even a single byte of slop after result.ptr+result.size, but at that point, for the sake of simplicity, we may be better off with the original implementation of
strbuf_attach()that automatically does that for us.
It seems reasonable to stay with the original implementation. So I'll change this patch to keep original calls to strbuf_attach() in rerere and apply. And I'll remove patch which enforces strbuf_attach() contract with BUG().
v5 cover http: add support for HTTP 429 rate limit retries
In reply to Vaidas Pilkauskas via GitGitGadgetChanges since v4:
* fix only strbuf_attach() calls which don't need reallocation
* remove patch, which enforces strbuf_attach() contract via BUG()
Changes since v3:
* Clean up of all strbuf_attach() call sites
* Add strbuf_attach() contract enforcement via BUG()
Changes since v2:
* New preparatory patch: Introduced show_http_message_fatal() helper function to reduce code duplication in remote-curl.c (suggested by Taylor Blau)
* Removed specific HTTP_RATE_LIMITED error handling from http-push.c and http-walker.c for the obsolete "dumb" protocol, allowing generic error handling to take over (suggested by Jeff King)
* Added support for CURLINFO_RETRY_AFTER on curl >= 7.66.0, falling back to manual header parsing on older versions
* Simplified retry/delay architecture: replaced complex non-blocking "delayed slot" mechanism with simple blocking sleep() call in the retry loop, removing ~66 lines of timing logic (suggested by Jeff King)
* Fixed Retry-After: 0 handling to allow immediate retry as specified by RFC 9110
* Changed http.retryAfter default from -1 to 0, so Git will retry immediately when encountering HTTP 429 without a Retry-After header, rather than failing with a configuration error
* Improved error messages: shortened to be more concise
* Fixed coding style issues: removed unnecessary curly braces, changed x == 0 to !x (per CodingGuidelines)
* Improved test portability: replaced non-portable date(1) commands with test-tool date, added nanosecond-precision timing with getnanos, replaced cut(1) with POSIX shell parameter expansion
* Split out strbuf.c bugfix into separate preparatory patch (the strbuf_reencode alloc size fix is unrelated to HTTP 429 support)
* Squashed separate trace2 logging patch into main HTTP 429 retry support commit
* Kept header_is_last_match assignment for Retry-After to prevent incorrect handling of HTTP header continuation lines
The implementation includes:
1. A bug fix in strbuf_reencode() that corrects the allocation size passed to strbuf_attach(), passing len+1 instead of len so that the existing buffer is reused rather than immediately reallocated.
2. A cleanup of strbuf_attach() call sites that were passing alloc == len, leaving no room for the NUL terminator. Sites with a known-NUL-terminated buffer now pass len+1; sites where the source buffer has no trailing NUL (ll_merge output) are converted to use strbuf_add() instead.
3. A new show_http_message_fatal() helper in remote-curl.c that combines the repeated pattern of show_http_message() followed by die() into a single NORETURN function, reducing boilerplate at existing call sites and providing a clean hook for the retry logic.
4. The main feature: HTTP 429 retry logic with support for the Retry-After header (both delay-seconds and HTTP-date formats), configurable via http.maxRetries, http.retryAfter, and http.maxRetryTime options. If any computed delay exceeds maxRetryTime the request fails immediately with a clear diagnostic rather than capping and retrying silently.
Vaidas Pilkauskas (4):
strbuf: pass correct alloc to strbuf_attach() in strbuf_reencode()
strbuf_attach: fix call sites to pass correct alloc
remote-curl: introduce show_http_message_fatal() helper
http: add support for HTTP 429 rate limit retries
Documentation/config/http.adoc | 23 +++ builtin/am.c | 2 +- builtin/fast-import.c | 2 +- git-curl-compat.h | 8 + http.c | 190 +++++++++++++++++++++-- http.h | 2 + mailinfo.c | 2 +- refs/files-backend.c | 2 +- remote-curl.c | 49 +++--- strbuf.c | 2 +- t/lib-httpd.sh | 1 + t/lib-httpd/apache.conf | 8 + t/lib-httpd/http-429.sh | 98 ++++++++++++ t/meson.build | 1 + t/t5584-http-429-retry.sh | 266 +++++++++++++++++++++++++++++++++ trailer.c | 2 +- 16 files changed, 623 insertions(+), 35 deletions(-) create mode 100644 t/lib-httpd/http-429.sh create mode 100755 t/t5584-http-429-retry.sh
base-commit: 7c02d39fc2ed2702223c7674f73150d9a7e61ba4
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-2008%2Fvaidas-shopify%2Fretry-after-v5
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-2008/vaidas-shopify/retry-after-v5
Pull-Request: https://github.com/gitgitgadget/git/pull/2008
Range-diff vs v4:
1: a3386f5b56 = 1: 7ec2d66447 strbuf: pass correct alloc to strbuf_attach() in strbuf_reencode() 2: f48b1f07c4 ! 2: 3e0b78cfb6 strbuf_attach: fix all call sites to pass correct alloc
@@ Metadata
Author: Vaidas Pilkauskas <[email protected]>
## Commit message ##
- strbuf_attach: fix all call sites to pass correct alloc
+ strbuf_attach: fix call sites to pass correct alloc
strbuf_attach(sb, buf, len, alloc) requires alloc > len (the buffer must have at least len+1 bytes to hold the NUL). Several call sites passed alloc == len, relying on strbuf_grow(sb, 0) inside strbuf_attach
- to reallocate. Prepare for changing that by fixing call sites to pass
- the correct alloc.
-
- - mailinfo, am, refs/files-backend, fast-import, trailer: pass len+1
- when the buffer is a NUL-terminated string (or from strbuf_detach).
- - rerere, apply: ll_merge returns a buffer with exactly result.size
- bytes (no extra NUL). Use strbuf_add() to copy and NUL-terminate
- into the strbuf, then free the merge result, so alloc is correct.
+ to reallocate. Fix these in mailinfo, am, refs/files-backend,
+ fast-import, and trailer by passing len+1 when the buffer is a
+ NUL-terminated string (or from strbuf_detach).
- ## apply.c ##
-@@ apply.c: static int three_way_merge(struct apply_state *state,
- return -1;
- }
- image_clear(image);
-- strbuf_attach(&image->buf, result.ptr, result.size, result.size);
-+ strbuf_add(&image->buf, result.ptr, result.size);
-+ free(result.ptr);
-
- return status;
- }
-
## builtin/am.c ##
@@ builtin/am.c: static void am_append_signoff(struct am_state *state)
{
@@ refs/files-backend.c: static int commit_ref(struct ref_lock *lock)
/*
* If this fails, commit_lock_file() will also fail
- ## rerere.c ##
-@@ rerere.c: static int handle_cache(struct index_state *istate,
- else
- io.io.output = NULL;
- strbuf_init(&io.input, 0);
-- strbuf_attach(&io.input, result.ptr, result.size, result.size);
-+ strbuf_add(&io.input, result.ptr, result.size);
-+ free(result.ptr);
-
- /*
- * Grab the conflict ID and optionally write the original
-
## trailer.c ##
@@ trailer.c: static struct trailer_block *trailer_block_get(const struct process_trailer_opti
for (ptr = trailer_lines; *ptr; ptr++) {
3: 557fd77444 < -: ---------- strbuf: replace strbuf_grow() in strbuf_attach() with BUG() check 4: 3a39dc9e39 = 3: 973703e9dd remote-curl: introduce show_http_message_fatal() helper 5: 5e0f4a56ef = 4: bfee1f10c0 http: add support for HTTP 429 rate limit retries
Signature · 1 line
gitgitgadget
PATCH v5 1/4 strbuf: pass correct alloc to strbuf_attach() in strbuf_reencode()
From: Vaidas Pilkauskas <[email protected]>
reencode_string_len() allocates len+1 bytes (including the NUL) and returns the string length in len. strbuf_reencode() was calling strbuf_attach(sb, out, len, len), so alloc was one byte too small.
strbuf_attach() then calls strbuf_grow(sb, 0). With alloc < len+1, ALLOC_GROW always reallocates, so we reallocated immediately after attach even when the strbuf was not extended further. Pass len+1 as the alloc argument so the existing buffer is reused and the reallocation is avoided.
--- strbuf.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
strbuf.c +1 -1
@@ -168,7 +168,7 @@ int strbuf_reencode(struct strbuf *sb, const char *from, const char *to)if (!out)return -1 ; -strbuf_attach (sb, out, len, len); +strbuf_attach (sb, out, len, len +1 );return 0 ; }
Signature · 1 line
gitgitgadget
PATCH v5 2/4 strbuf_attach: fix call sites to pass correct alloc
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
strbuf_attach(sb, buf, len, alloc) requires alloc > len (the buffer must have at least len+1 bytes to hold the NUL). Several call sites passed alloc == len, relying on strbuf_grow(sb, 0) inside strbuf_attach to reallocate. Fix these in mailinfo, am, refs/files-backend, fast-import, and trailer by passing len+1 when the buffer is a NUL-terminated string (or from strbuf_detach).
--- builtin/am.c | 2 +- builtin/fast-import.c | 2 +- mailinfo.c | 2 +- refs/files-backend.c | 2 +- trailer.c | 2 +- 5 files changed, 5 insertions(+), 5 deletions(-)
builtin/am.c +1 -1
@@ -1188,7 +1188,7 @@ static void am_append_signoff(struct am_state *state) {struct strbuf sb =STRBUF_INIT ; -strbuf_attach (&sb, state->msg, state->msg_len, state->msg_len); +strbuf_attach (&sb, state->msg, state->msg_len, state->msg_len +1 );append_signoff (&sb,0 ,0 ); state->msg =strbuf_detach (&sb, &state->msg_len); }
builtin/fast-import.c +1 -1
@@ -3246,7 +3246,7 @@ static void cat_blob(struct object_entry *oe, struct object_id *oid)cat_blob_write ("\n" ,1 );if (oe && oe->pack_id == pack_id) { last_blob.offset = oe->idx.offset; -strbuf_attach (&last_blob.data, buf, size, size); +strbuf_attach (&last_blob.data, buf, size, size +1 ); last_blob.depth = oe->depth; }else free (buf);
mailinfo.c +1 -1
@@ -470,7 +470,7 @@ static int convert_to_utf8(struct mailinfo *mi,return error ("cannot convert from %s to %s" , charset, mi->metainfo_charset); } -strbuf_attach (line, out, out_len, out_len); +strbuf_attach (line, out, out_len, out_len +1 );return 0 ; }
refs/files-backend.c +1 -1
@@ -1806,7 +1806,7 @@ static int commit_ref(struct ref_lock *lock)size_t len =strlen (path);struct strbuf sb_path =STRBUF_INIT ; -strbuf_attach (&sb_path, path, len, len); +strbuf_attach (&sb_path, path, len, len +1 );/* * If this fails, commit_lock_file() will also fail
trailer.c +1 -1
@@ -1009,7 +1009,7 @@ static struct trailer_block *trailer_block_get(const struct process_trailer_optifor (ptr = trailer_lines; *ptr; ptr++) {if (last &&isspace ((*ptr)->buf[0 ])) {struct strbuf sb =STRBUF_INIT ; -strbuf_attach (&sb, *last,strlen (*last),strlen (*last)); +strbuf_attach (&sb, *last,strlen (*last),strlen (*last) +1 );strbuf_addbuf (&sb, *ptr); *last =strbuf_detach (&sb,NULL );continue ;
Signature · 1 line
gitgitgadget
PATCH v5 3/4 remote-curl: introduce show_http_message_fatal() helper
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
Several code paths in remote-curl.c follow the same pattern of calling show_http_message() to display server error messages followed by die() to terminate with an error. This duplication makes the code more verbose and harder to maintain.
Introduce a new show_http_message_fatal() helper function that combines these two operations. This function:
1. Displays any HTTP error message from the server via show_http_message()
2. Calls die() with the provided error message
3. Returns NORETURN to help the compiler with control flow analysis
Refactor existing call sites in remote-curl.c to use this new helper, reducing code duplication and improving readability. This pattern will also be used by upcoming HTTP 429 rate limiting support.
--- remote-curl.c | 45 ++++++++++++++++++++++++++++----------------- 1 file changed, 28 insertions(+), 17 deletions(-)
remote-curl.c +28 -17
@@ -367,23 +367,25 @@ static void free_discovery(struct discovery *d) } } -static int show_http_message (struct strbuf *type,struct strbuf *charset, -struct strbuf *msg) +static NORETURN void show_http_message_fatal(struct strbuf *type, struct strbuf *charset, +struct strbuf *msg,const char *fmt, ...) {const char *p, *eol; +va_list ap; +report_fn die_message_routine =get_die_message_routine ();/* * We only show text/plain parts, as other types are likely * to be ugly to look at on the user's terminal. */ if (strcmp (type->buf,"text/plain" )) -return -1 ; + goto out;if (charset->len)strbuf_reencode (msg, charset->buf,get_log_output_encoding ());strbuf_trim (msg);if (!msg->len) -return -1 ; + goto out; p = msg->buf; do {
@@ -391,7 +393,16 @@ static int show_http_message(struct strbuf *type, struct strbuf *charset,fprintf (stderr,"remote: %.*s\n" , (int )(eol - p), p); p = eol +1 ; }while (*eol); -return 0 ; + +out: +strbuf_release (type); +strbuf_release (charset); +strbuf_release (msg); + +va_start (ap, fmt); +die_message_routine (fmt, ap); +va_end (ap); +exit (128 ); } static int get_protocol_http_header(enum protocol_version version,
@@ -518,21 +529,21 @@ static struct discovery *discover_refs(const char *service, int for_push)case HTTP_OK :break ;case HTTP_MISSING_TARGET : -show_http_message (&type, &charset, &buffer); -die (_ ("repository '%s' not found" ), -transport_anonymize_url (url.buf)); +show_http_message_fatal (&type, &charset, &buffer, +_ ("repository '%s' not found" ), +transport_anonymize_url (url.buf));case HTTP_NOAUTH : -show_http_message (&type, &charset, &buffer); -die (_ ("Authentication failed for '%s'" ), -transport_anonymize_url (url.buf)); +show_http_message_fatal (&type, &charset, &buffer, +_ ("Authentication failed for '%s'" ), +transport_anonymize_url (url.buf));case HTTP_NOMATCHPUBLICKEY : -show_http_message (&type, &charset, &buffer); -die (_ ("unable to access '%s' with http.pinnedPubkey configuration: %s" ), -transport_anonymize_url (url.buf), curl_errorstr); +show_http_message_fatal (&type, &charset, &buffer, +_ ("unable to access '%s' with http.pinnedPubkey configuration: %s" ), +transport_anonymize_url (url.buf), curl_errorstr);default : -show_http_message (&type, &charset, &buffer); -die (_ ("unable to access '%s': %s" ), -transport_anonymize_url (url.buf), curl_errorstr); +show_http_message_fatal (&type, &charset, &buffer, +_ ("unable to access '%s': %s" ), +transport_anonymize_url (url.buf), curl_errorstr); }if (options.verbosity && !starts_with (refs_url.buf, url.buf)) {
Signature · 1 line
gitgitgadget
PATCH v5 4/4 http: add support for HTTP 429 rate limit retries
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
Add retry logic for HTTP 429 (Too Many Requests) responses to handle server-side rate limiting gracefully. When Git's HTTP client receives a 429 response, it can now automatically retry the request after an appropriate delay, respecting the server's rate limits.
The implementation supports the RFC-compliant Retry-After header in both delay-seconds (integer) and HTTP-date (RFC 2822) formats. If a past date is provided, Git retries immediately without waiting.
Retry behavior is controlled by three new configuration options (http.maxRetries, http.retryAfter, and http.maxRetryTime) which are documented in git-config(1).
The retry logic implements a fail-fast approach: if any delay (whether from server header or configuration) exceeds maxRetryTime, Git fails immediately with a clear error message rather than capping the delay. This provides better visibility into rate limiting issues.
The implementation includes extensive test coverage for basic retry behavior, Retry-After header formats (integer and HTTP-date), configuration combinations, maxRetryTime limits, invalid header handling, environment variable overrides, and edge cases.
--- Documentation/config/http.adoc | 23 +++ git-curl-compat.h | 8 + http.c | 190 +++++++++++++++++++++-- http.h | 2 + remote-curl.c | 4 + t/lib-httpd.sh | 1 + t/lib-httpd/apache.conf | 8 + t/lib-httpd/http-429.sh | 98 ++++++++++++ t/meson.build | 1 + t/t5584-http-429-retry.sh | 266 +++++++++++++++++++++++++++++++++ 10 files changed, 589 insertions(+), 12 deletions(-) create mode 100644 t/lib-httpd/http-429.sh create mode 100755 t/t5584-http-429-retry.sh
Documentation/config/http.adoc +23 -0
@@ -315,6 +315,29 @@ http.keepAliveCount:: unset, curl's default value is used. Can be overridden by the `GIT_HTTP_KEEPALIVE_COUNT` environment variable. +http.retryAfter:: + Default wait time in seconds before retrying when a server returns + HTTP 429 (Too Many Requests) without a Retry-After header. + Defaults to 0 (retry immediately). When a Retry-After header is + present, its value takes precedence over this setting. Can be + overridden by the `GIT_HTTP_RETRY_AFTER` environment variable. + See also `http.maxRetries` and `http.maxRetryTime`. + +http.maxRetries:: + Maximum number of times to retry after receiving HTTP 429 (Too Many + Requests) responses. Set to 0 (the default) to disable retries. + Can be overridden by the `GIT_HTTP_MAX_RETRIES` environment variable. + See also `http.retryAfter` and `http.maxRetryTime`. + +http.maxRetryTime:: + Maximum time in seconds to wait for a single retry attempt when + handling HTTP 429 (Too Many Requests) responses. If the server + requests a delay (via Retry-After header) or if `http.retryAfter` + is configured with a value that exceeds this maximum, Git will fail + immediately rather than waiting. Default is 300 seconds (5 minutes). + Can be overridden by the `GIT_HTTP_MAX_RETRY_TIME` environment + variable. See also `http.retryAfter` and `http.maxRetries`. + http.noEPSV:: A boolean which disables using of EPSV ftp command by curl. This can be helpful with some "poor" ftp servers which don't
git-curl-compat.h +8 -0
@@ -37,6 +37,14 @@#define GIT_CURL_NEED_TRANSFER_ENCODING_HEADER #endif +/** +* CURLINFO_RETRY_AFTER was added in 7.66.0, released in September 2019. +* It allows curl to automatically parse Retry-After headers. +*/ +#if LIBCURL_VERSION_NUM >=0x074200 +#define GIT_CURL_HAVE_CURLINFO_RETRY_AFTER 1 +#endif +/** * CURLOPT_PROTOCOLS_STR and CURLOPT_REDIR_PROTOCOLS_STR were added in 7.85.0, * released in August 2022.
http.c +178 -12
305 lines not on this page · raw message
http.h +2 -0
@@ -20,6 +20,7 @@ struct slot_results {long http_code;long auth_avail;long http_connectcode; +long retry_after; };struct active_request_slot {
@@ -167,6 +168,7 @@ struct http_get_options {#define HTTP_REAUTH 4#define HTTP_NOAUTH 5#define HTTP_NOMATCHPUBLICKEY 6 +#define HTTP_RATE_LIMITED 7/* * Requests a URL and stores the result in a strbuf.
remote-curl.c +4 -0
@@ -540,6 +540,10 @@ static struct discovery *discover_refs(const char *service, int for_push)show_http_message_fatal (&type, &charset, &buffer,_ ("unable to access '%s' with http.pinnedPubkey configuration: %s" ),transport_anonymize_url (url.buf), curl_errorstr); +case HTTP_RATE_LIMITED : +show_http_message_fatal (&type, &charset, &buffer, +_ ("rate limited by '%s', please try again later" ), +transport_anonymize_url (url.buf));default : show_http_message_fatal(&type, &charset, &buffer, _("unable to access '%s': %s"),
t/lib-httpd.sh +1 -0
@@ -167,6 +167,7 @@ prepare_httpd() { install_script error.sh install_script apply-one-time-script.sh install_script nph-custom-auth.sh + install_script http-429.sh ln -s "$LIB_HTTPD_MODULE_PATH" "$HTTPD_ROOT_PATH/modules"
t/lib-httpd/apache.conf +8 -0
@@ -139,6 +139,10 @@ SetEnv PERL_PATH ${PERL_PATH} SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} SetEnv GIT_HTTP_EXPORT_ALL </LocationMatch> +<LocationMatch /http_429/> + SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} + SetEnv GIT_HTTP_EXPORT_ALL +</LocationMatch> <LocationMatch /smart_v0/> SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} SetEnv GIT_HTTP_EXPORT_ALL
@@ -160,6 +164,7 @@ ScriptAlias /broken_smart/ broken-smart-http.sh/ ScriptAlias /error_smart/ error-smart-http.sh/ ScriptAlias /error/ error.sh/ ScriptAliasMatch /one_time_script/(.*) apply-one-time-script.sh/$1 +ScriptAliasMatch /http_429/(.*) http-429.sh/$1 ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1 <Directory ${GIT_EXEC_PATH}> Options FollowSymlinks
@@ -185,6 +190,9 @@ ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1 <Files apply-one-time-script.sh> Options ExecCGI </Files> +<Files http-429.sh> + Options ExecCGI +</Files> <Files ${GIT_EXEC_PATH}/git-http-backend> Options ExecCGI </Files>
t/lib-httpd/http-429.sh new file +98 -0
98 lines not on this page · raw message
t/meson.build +1 -0
@@ -704,6 +704,7 @@ integration_tests = [ 't5581-http-curl-verbose.sh', 't5582-fetch-negative-refspec.sh', 't5583-push-branches.sh', + 't5584-http-429-retry.sh', 't5600-clone-fail-cleanup.sh', 't5601-clone.sh', 't5602-clone-remote-exec.sh',
t/t5584-http-429-retry.sh new file +266 -0
266 lines not on this page · raw message
Signature · 1 line
gitgitgadget
"Vaidas Pilkauskas via GitGitGadget" <[email protected]>
Quoted 9 lines · writes:
Changes since v4:
* fix only
strbuf_attach()calls which don't need reallocation
* remove patch, which enforcesstrbuf_attach()contract viaBUG()
...
Vaidas Pilkauskas (4):
strbuf: pass correct alloc tostrbuf_attach()instrbuf_reencode()
strbuf_attach: fix call sites to pass correct alloc
remote-curl: introduceshow_http_message_fatal()helper
These three patches looked quite reasonable to me.
Quoted 1 line
http: add support for HTTP 429 rate limit retries
I'd feel comfortable to see somebody more familiar with the HTTP transport code base to take a look at this step before we declare victory.
Thanks.
Quoted 13 lines · Junio C Hamano
"Vaidas Pilkauskas via GitGitGadget" <[email protected]>
Quoted 9 lines · writes:
Changes since v4:
* fix only
strbuf_attach()calls which don't need reallocation
* remove patch, which enforcesstrbuf_attach()contract viaBUG()
...
Vaidas Pilkauskas (4):
strbuf: pass correct alloc tostrbuf_attach()instrbuf_reencode()
strbuf_attach: fix call sites to pass correct alloc
remote-curl: introduceshow_http_message_fatal()helper
Quoted 7 lines
These three patches looked quite reasonable to me.
Quoted 1 line
http: add support for HTTP 429 rate limit retries
I'd feel comfortable to see somebody more familiar with the HTTP transport code base to take a look at this step before we declare victory.
Any volunteers?
Thanks.
Re: [PATCH v5 3/4] remote-curl: introduce show_http_message_fatal() helper
In reply to Vaidas Pilkauskas via GitGitGadgetQuoted 10 lines · Vaidas Pilkauskas via GitGitGadget
From: Vaidas Pilkauskas <[email protected]>
Several code paths in
remote-curl.cfollow the same pattern of callingshow_http_message()to display server error messages followed bydie()to terminate with an error. This duplication makes the code more verbose and harder to maintain.Introduce a new
show_http_message_fatal()helper function that combines these two operations. This function:
Quoted 7 lines
1. Displays any HTTP error message from the server via
show_http_message()
2. Callsdie()with the provided error message
3. Returns NORETURN to help the compiler with control flow analysisRefactor existing call sites in
remote-curl.cto use this new helper, reducing code duplication and improving readability. This pattern will also be used by upcoming HTTP 429 rate limiting support.
I'm unconvinced that this actually makes things simpler. It doesn't save any lines, the show_http_message() becomes less flexible, and we now encode die-logic like the numeric code for exit() in the curl code.
remote-curl.c static void free_discovery(struct discovery *d) · line 375
Quoted 6 lines
remote-curl.c +4
-0+static NORETURN void show_http_message_fatal(struct strbuf *type, struct strbuf *charset, +struct strbuf *msg,const char *fmt, ...) {const char *p, *eol; +va_list ap; +report_fn die_message_routine =get_die_message_routine ();
If we got the die_routine here and not the die_message_routine, we wouldn't have to exit() ourselves. But sadly there is not a get_die_routine() helper yet, so we'd have to add one.
Quoted 7 lines
remote-curl.c +1
-1/* * We only show text/plain parts, as other types are likely * to be ugly to look at on the user's terminal. */ if (strcmp (type->buf,"text/plain" )) -return -1 ; + goto out;
remote-curl.c static void free_discovery(struct discovery *d) · line 391
Quoted 10 lines
remote-curl.c +1
-1if (charset->len)strbuf_reencode (msg, charset->buf,get_log_output_encoding ());strbuf_trim (msg);if (!msg->len) -return -1 ; + goto out; p = msg->buf; do {
OK, we jump to the end since now we have to make sure to hit the exit bits.
remote-curl.c static int show_http_message(struct strbuf *type, struct strbuf *charset, · line 400
Quoted 10 lines
remote-curl.c +5
-1@@ -391,7 +393,16 @@ static int show_http_message(struct strbuf *type, struct strbuf *charset,fprintf (stderr,"remote: %.*s\n" , (int )(eol - p), p); p = eol +1 ; }while (*eol); -return 0 ; + +out: +strbuf_release (type); +strbuf_release (charset); +strbuf_release (msg);
This cleanup seems pointless. These strbufs came from the caller, so they are not our responsibility to release. But also, we're about to call die(), so there is no need to clean them up.
remote-curl.c static int show_http_message(struct strbuf *type, struct strbuf *charset, · line 405
Quoted 4 lines
remote-curl.c +4
-0+va_start (ap, fmt); +die_message_routine (fmt, ap); +va_end (ap); +exit (128 );
This part is the advertised change for the function, which looks correct.
remote-curl.c static struct discovery *discover_refs(const char *service, int for_push) · line 534
Quoted 11 lines
remote-curl.c +3
-3static int get_protocol_http_header(enum protocol_version version,@@ -518,21 +529,21 @@ static struct discovery *discover_refs(const char *service, int for_push)case HTTP_OK :break ;case HTTP_MISSING_TARGET : -show_http_message (&type, &charset, &buffer); -die (_ ("repository '%s' not found" ), -transport_anonymize_url (url.buf)); +show_http_message_fatal (&type, &charset, &buffer, +_ ("repository '%s' not found" ), +transport_anonymize_url (url.buf));
So this is what the refactoring bought us: we trade 3 lines for 3 lines. ;)
If we could roll in policy decisions like anonymizing the URL, or showing the curl_errorstr, I think it might be worth it. But we can't do that easily because they are varargs for the %s format.
IMHO we should just drop this patch, and patch 4 should do the usual show_http_message() and die().
-Peff
Quoted 41 lines · Vaidas Pilkauskas via GitGitGadget
(quoted diff) +37
-0@@ -257,6 +264,50 @@ static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UN goto exit; } +#ifndef GIT_CURL_HAVE_CURLINFO_RETRY_AFTER + /* Parse Retry-After header for rate limiting (for curl < 7.66.0) */ + if (skip_iprefix_mem(ptr, size, "retry-after:", &val, &val_len)) { + struct active_request_slot *slot = (struct active_request_slot *)p; + + strbuf_add(&buf, val, val_len); + strbuf_trim(&buf); + + if (slot && slot->results) { + /* Parse the retry-after value (delay-seconds or HTTP-date) */ + char *endptr; + long retry_after; + + errno = 0; + retry_after = strtol(buf.buf, &endptr, 10); + + /* Check if it's a valid integer (delay-seconds format) */ + if (endptr != buf.buf && *endptr == '\0' && + errno != ERANGE && retry_after >= 0) { + slot->results->retry_after = retry_after; + } else { + /* Try parsing as HTTP-date format */ + timestamp_t timestamp; + int offset; + if (!parse_date_basic(buf.buf, ×tamp, &offset)) { + /* Successfully parsed as date, calculate delay from now */ + timestamp_t now = time(NULL); + if (timestamp > now) { + slot->results->retry_after = (long)(timestamp - now); + } else { + /* Past date means retry immediately */ + slot->results->retry_after = 0; + } + } else { + /* Failed to parse as either delay-seconds or HTTP-date */ + warning(_("unable to parse Retry-After header value: '%s'"), buf.buf); + }
Review static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UN · line 313
Quoted 10 lines
(quoted diff) +7
-0+ } + } + + goto exit; + } +#endif + /* * This line could be a continuation of the previously matched header * field. If this is the case then we should append this value to the
I don't think this meshes well with the existing code for parsing www-authenticate, especially with respect to header continuation:
1. If we see continuation like "Retry-After:\n <date>", we won't find <date>. Instead, we'll just think it's blank (or worse, do a partial parse if the line break happens at whitespace in the middle of the date).
2. We don't reset http_auth.header_is_last_match, so "WWW-Authenticate: foo\nRetry-After:\n <date>" will attribute <date> to the WWW-Authenticate header.
3. We don't clear the value like we do for www-authenticate headers, as we do in the lower code:
/*
* If this is a HTTP status line and not a header field, this signals
* a different HTTP response. libcurl writes all the output of all
* response headers of all responses, including redirects.
* We only care about the last HTTP request response's headers so clear
* the existing array.
*/
if (skip_iprefix_mem(ptr, size, "http/", &val, &val_len))
strvec_clear(values);
It's probably unlikely to get a retry-after _and_ a redirect in practice, though.
I suspect these are all quite uncommon cases, but it worries me a little that a malicious response could inject values into the wrong header. And worse, that these cases will go largely untested, as most developers have recent enough versions of libcurl.
Could we drop this hunk entirely, and just document that the Retry-After feature only works if you have a recent version of libcurl? Curl 7.66.0 is almost 7 years old now (and you'd still get manual-timed retries if you configure them).
Quoted 5 lines
(quoted diff) +1
-0@@ -342,6 +393,17 @@ static void finish_active_slot(struct active_request_slot *slot) curl_easy_getinfo(slot->curl, CURLINFO_HTTP_CONNECTCODE, &slot->results->http_connectcode); +
Review static void finish_active_slot(struct active_request_slot *slot) · line 406
Quoted 10 lines
(quoted diff) +10
-0+#ifdef GIT_CURL_HAVE_CURLINFO_RETRY_AFTER + if (slot->results->http_code == 429) { + curl_off_t retry_after; + CURLcode res = curl_easy_getinfo(slot->curl, + CURLINFO_RETRY_AFTER, + &retry_after); + if (res == CURLE_OK && retry_after > 0) + slot->results->retry_after = (long)retry_after; + } +#endif
This hunk makes sense. It is a little funny that we cast to long after getting a curl_off_t, but I suspect we have to cast somewhere unless we want to consistently pass around a curl_off_t (including via http_request).
Though since it seems like http_request() is the only code path that handles retries, do we need to grab it here in finish_active_slot()? I.e., would it make more sense to do it in http_request() where we are copying it out from the results field anyway?
Pushing it down to this level would make sense if we wanted to handle retries on other requests outside of http_request (like dumb-http walker request), but I don't think your patch does that.
Review int run_one_slot(struct active_request_slot *slot, · line 1976
Quoted 7 lines
(quoted diff) +3
-0@@ -1886,6 +1971,9 @@ int run_one_slot(struct active_request_slot *slot, struct slot_results *results) { slot->results = results; + /* Initialize retry_after to -1 (not set) */ + results->retry_after = -1; +
This is an unusual spot for per-request setup. Usually this happens in get_active_slot(), which is called before making a request.
I think you are putting it here because we will make several requests on the same handle via http_request_recoverable(). But in that case, would setting it there make more sense? In fact, we seem to _reset_ it there already.
Review int run_one_slot(struct active_request_slot *slot,
Quoted 6 lines
(quoted diff) +2
-1static int http_request(const char *url, void *result, int target, - const struct http_get_options *options) + const struct http_get_options *options, + long *retry_after_out) {
Why add this as a new argument when we already have many other optional out-parameters in http_get_options? I.e., I'd have expected this:
http.h +2 -0
@@ -155,12 +155,14 @@ struct http_get_options {/* * If not NULL, contains additional HTTP headers to be sent with the * request. The strings in the list must not be freed until after the * request has completed. */ struct string_list *extra_headers; + +long *retry_after; };/* Return values for http_get_*() */ #define HTTP_OK 0#define HTTP_MISSING_TARGET 1#define HTTP_ERROR 2
We could possibly even just make it a regular "long", not a pointer. The reason the other fields in http_get_options are pointers is that we only want to fill them in if the caller is interested, because:
1. There is a cost to getting/computing the information.
2. They allocate memory, so the caller has to know to clean them up.
But here we just have a plain integer, so we could just always assign it. It would mean dropping the "const" from the struct parameter, but I think that's fine.
Looking further in the patch, I guess one reason is that the retry stuff is handled internally by http_request(), so a caller never needs or wants to see it. I wonder if that is always true, though. For example, when we print the rate-limited message at the top-level of remote-curl.c, might it be useful for us to mention the retry-after field?
If you want retries to work for all calls, we might have to use a dummy http_get_options struct. But I think that would actually make the code cleaner in general (no more "if (options && options->effective_url)"; it would just "if (options->effective_url)".
Review static int http_request(const char *url, · line 2241
Quoted 7 lines
(quoted diff) +2
-1@@ -2148,7 +2237,8 @@ static int http_request(const char *url, fwrite_buffer); } - curl_easy_setopt(slot->curl, CURLOPT_HEADERFUNCTION, fwrite_wwwauth); + curl_easy_setopt(slot->curl, CURLOPT_HEADERFUNCTION, fwrite_headers); + curl_easy_setopt(slot->curl, CURLOPT_HEADERDATA, slot);
This hunk goes away if we drop the fwrite_headers() bit.
Quoted 5 lines
(quoted diff) +4
-0+static long handle_rate_limit_retry(int *rate_limit_retries, long slot_retry_after) + /* Use the slot-specific retry_after value or configured default */ + if (slot_retry_after >= 0) { + /* Check if retry delay exceeds maximum allowed */
Review static int http_request(const char *url,
Quoted 10 lines
(quoted diff) +10
-0+ if (slot_retry_after > http_max_retry_time) { + error(_("response requested a delay greater than http.maxRetryTime (%ld > %ld seconds)"), + slot_retry_after, http_max_retry_time); + trace2_data_string("http", the_repository, + "http/429-error", "exceeds-max-retry-time"); + trace2_data_intmax("http", the_repository, + "http/429-requested-delay", slot_retry_after); + return -1; + } + return slot_retry_after;
OK. I had imagine that the max-time would clamp how long we were willing to wait, but instead we just bail. I guess what you have is probably more friendly to a server that says "please try again after 600 seconds" as opposed to trying again faster than they'd prefer.
Review
Quoted 9 lines
(quoted diff) +9
-0+ } else { + /* No Retry-After header provided, use configured default */ + if (http_retry_after > http_max_retry_time) { + error(_("configured http.retryAfter exceeds http.maxRetryTime (%ld > %ld seconds)"), + http_retry_after, http_max_retry_time); + trace2_data_string("http", the_repository, + "http/429-error", "config-exceeds-max-retry-time"); + return -1; + }
And this one is a misconfiguration on the part of the user, so it should certainly be an error. ;)
Review
Quoted 8 lines
(quoted diff) +3
-0+static int http_request_recoverable(const char *url, void *result, int target, struct http_get_options *options) { int i = 3; int ret; + int rate_limit_retries = http_max_retries; + long slot_retry_after = -1; /* Per-slot retry_after value */
Hmm, what is "i" here? Previously we used it to avoid REAUTH looping too many times:
Review
Quoted 2 lines
(quoted diff) +1
-1- while (ret == HTTP_REAUTH && --i) { + while ((ret == HTTP_REAUTH || ret == HTTP_RATE_LIMITED) && --i) {
But now we are looping on both rate-limits and auth. If max_retries is greater than 2, won't we bail even though there are retries left?
Also, what if we retry once due to RATE_LIMITED, but then need multiple attempts to do auth? It feels like we should have two separate counters:
while ((ret == HTTP_REAUTH && --i) ||
(ret == HTTP_RATE_LIMITED && --rate_limit_retries)Side note: I think I am reading the "--i" count right. It feels like it would be simpler as "i--", but I guess it is counting the initial http_request() we made. I suspect this might make more sense as a do-while loop, but it is hairy enough that trying to refactor it further might be risky.
Quoted 9 lines
(quoted diff) +5
-0@@ -2301,10 +2454,23 @@ static int http_request_reauth(const char *url, default: BUG("Unknown http_request target"); } + if (ret == HTTP_RATE_LIMITED) { + retry_delay = handle_rate_limit_retry(&rate_limit_retries, slot_retry_after); + if (retry_delay < 0) + return HTTP_ERROR; +
Review static int http_request_reauth(const char *url, · line 2471
Quoted 10 lines
(quoted diff) +10
-0+ if (retry_delay > 0) { + warning(_("rate limited, waiting %ld seconds before retry"), retry_delay); + trace2_data_intmax("http", the_repository, + "http/retry-sleep-seconds", retry_delay); + sleep(retry_delay); + } + slot_retry_after = -1; /* Reset after use */ + } else if (ret == HTTP_REAUTH) { + credential_fill(the_repository, &http_auth, 1); + }
OK. In the loop condition I suggested above, we decrement rate_limit_retries there. But obviously it is also happening here via handle_rate_limit_retry(), so it has to be one or the other.
I wondered about the case where we are out of retries, and whether we might sleep before bailing. But that function checks that case and returns -1, so we'd bail immediately. Good.
Quoted 4 lines
(quoted diff) +4
-0+test_expect_success 'HTTP 429 retry delays are respected' ' + # Enable retries + test_config http.maxRetries 3 && +
Review static int http_request_reauth(const char *url,
Quoted 10 lines
(quoted diff) +10
-0+ # Time the operation - it should take at least 2 seconds due to retry delay + start=$(test-tool date getnanos) && + git ls-remote "$HTTPD_URL/http_429/retry-delays-respected/2/repo.git" >output 2>err && + duration=$(test-tool date getnanos $start) && + + # Verify it took at least 2 seconds (allowing some tolerance) + duration_int=${duration%.*} && + test "$duration_int" -ge 1 && + test_grep "refs/heads/" output +'
Nice, the duration-checking here looks nice and simple. The 2-second check here should be non-racy, since we'll wait at least that long. We may get a false-success on a heavily-loaded system, but that's OK.
Quoted 5 lines
(quoted diff) +5
-0+test_expect_success 'HTTP 429 fails immediately if Retry-After exceeds http.maxRetryTime' ' + # Configure max retry time to 3 seconds (much less than requested 100) + test_config http.maxRetries 3 && + test_config http.maxRetryTime 3 && +
Review
Quoted 10 lines
(quoted diff) +10
-0+ # Should fail immediately without waiting + start=$(test-tool date getnanos) && + test_must_fail git ls-remote "$HTTPD_URL/http_429/retry-after-exceeds-max-time/100/repo.git" 2>err && + duration=$(test-tool date getnanos $start) && + + # Should fail quickly (less than 2 seconds, no 100 second wait) + duration_int=${duration%.*} && + test "$duration_int" -lt 2 && + test_grep "greater than http.maxRetryTime" err +'
This one is the opposite, though. On a heavily loaded system, we may get a false failure if the test takes longer than 2s to run. Can we bump this to something much less likely to trigger, like 99? Or even 100 would work, I'd think, since we know we'll wait at least 100 seconds if we actually tried to use that retry-time.
Quoted 7 lines
(quoted diff) +7
-0+test_expect_success 'HTTP 429 fails if configured http.retryAfter exceeds http.maxRetryTime' ' + # Test misconfiguration: retryAfter > maxRetryTime + # Configure retryAfter larger than maxRetryTime + test_config http.maxRetries 3 && + test_config http.retryAfter 100 && + test_config http.maxRetryTime 5 && +
Review
Quoted 10 lines
(quoted diff) +10
-0+ # Should fail immediately with configuration error + start=$(test-tool date getnanos) && + test_must_fail git ls-remote "$HTTPD_URL/http_429/config-retry-after-exceeds-max-time/none/repo.git" 2>err && + duration=$(test-tool date getnanos $start) && + + # Should fail quickly + duration_int=${duration%.*} && + test "$duration_int" -lt 2 && + test_grep "configured http.retryAfter.*exceeds.*http.maxRetryTime" err +'
Same here; the key thing is that we don't wait 100 seconds, but we might accidentally take 2 seconds on a slow system.
I think there are a few more below, just looking for "-lt".
-Peff
Quoted 17 lines · Junio C Hamano
Quoted 15 lines · Junio C Hamano
"Vaidas Pilkauskas via GitGitGadget" <[email protected]>
Quoted 9 lines · writes:
Changes since v4:
* fix only
strbuf_attach()calls which don't need reallocation
* remove patch, which enforcesstrbuf_attach()contract viaBUG()
...
Vaidas Pilkauskas (4):
strbuf: pass correct alloc tostrbuf_attach()instrbuf_reencode()
strbuf_attach: fix call sites to pass correct alloc
remote-curl: introduceshow_http_message_fatal()helperThese three patches looked quite reasonable to me.
Quoted 7 lines
Quoted 5 lines
Quoted 1 line
http: add support for HTTP 429 rate limit retries
I'd feel comfortable to see somebody more familiar with the HTTP transport code base to take a look at this step before we declare victory.
Any volunteers?
Sorry, I'm way underwater on things I could/should be reviewing, and this one was quite non-trivial. ;)
I just left some comments. A lot of it was about how to more cleanly integrate with the http code (which I admit is a mess, especially with respect to which "layer" things should happen at). Some of that may be debatable, though I hope we can make things a bit cleaner.
But I think there may be a logic error in how http_request_recoverable() loops, which certainly needs to be fixed.
-Peff
Quoted 30 lines · Jeff King
Quoted 17 lines · Junio C Hamano
Quoted 15 lines · Junio C Hamano
"Vaidas Pilkauskas via GitGitGadget" <[email protected]>
Quoted 9 lines · writes:
Changes since v4:
* fix only
strbuf_attach()calls which don't need reallocation
* remove patch, which enforcesstrbuf_attach()contract viaBUG()
...
Vaidas Pilkauskas (4):
strbuf: pass correct alloc tostrbuf_attach()instrbuf_reencode()
strbuf_attach: fix call sites to pass correct alloc
remote-curl: introduceshow_http_message_fatal()helperThese three patches looked quite reasonable to me.
Quoted 7 lines
Quoted 5 lines
Quoted 1 line
http: add support for HTTP 429 rate limit retries
I'd feel comfortable to see somebody more familiar with the HTTP transport code base to take a look at this step before we declare victory.
Any volunteers?
Sorry, I'm way underwater on things I could/should be reviewing, and this one was quite non-trivial. ;)
Quoted 7 lines
I just left some comments. A lot of it was about how to more cleanly integrate with the http code (which I admit is a mess, especially with respect to which "layer" things should happen at). Some of that may be debatable, though I hope we can make things a bit cleaner.
But I think there may be a logic error in how
http_request_recoverable()loops, which certainly needs to be fixed.
Thanks.
v6 cover http: add support for HTTP 429 rate limit retries
In reply to Vaidas Pilkauskas via GitGitGadgetChanges since v5:
* drop show_http_message_fatal() preparation patch
* drop fwrite_headers, restore fwrite_wwwauth
* move CURLINFO_RETRY_AFTER from finish_active_slot to http_request
* move retry_after=-1 init from run_one_slot to http_request
* replace retry_after_out param with http_get_options field
* fix loop counter: separate REAUTH and RATE_LIMITED counters
* fix racy -lt 2 timing bounds in tests
Changes since v4:
* fix only strbuf_attach() calls which don't need reallocation
* remove patch, which enforces strbuf_attach() contract via BUG()
Changes since v3:
* Clean up of all strbuf_attach() call sites
* Add strbuf_attach() contract enforcement via BUG()
Changes since v2:
* New preparatory patch: Introduced show_http_message_fatal() helper function to reduce code duplication in remote-curl.c (suggested by Taylor Blau)
* Removed specific HTTP_RATE_LIMITED error handling from http-push.c and http-walker.c for the obsolete "dumb" protocol, allowing generic error handling to take over (suggested by Jeff King)
* Added support for CURLINFO_RETRY_AFTER on curl >= 7.66.0, falling back to manual header parsing on older versions
* Simplified retry/delay architecture: replaced complex non-blocking "delayed slot" mechanism with simple blocking sleep() call in the retry loop, removing ~66 lines of timing logic (suggested by Jeff King)
* Fixed Retry-After: 0 handling to allow immediate retry as specified by RFC 9110
* Changed http.retryAfter default from -1 to 0, so Git will retry immediately when encountering HTTP 429 without a Retry-After header, rather than failing with a configuration error
* Improved error messages: shortened to be more concise
* Fixed coding style issues: removed unnecessary curly braces, changed x == 0 to !x (per CodingGuidelines)
* Improved test portability: replaced non-portable date(1) commands with test-tool date, added nanosecond-precision timing with getnanos, replaced cut(1) with POSIX shell parameter expansion
* Split out strbuf.c bugfix into separate preparatory patch (the strbuf_reencode alloc size fix is unrelated to HTTP 429 support)
* Squashed separate trace2 logging patch into main HTTP 429 retry support commit
* Kept header_is_last_match assignment for Retry-After to prevent incorrect handling of HTTP header continuation lines
The implementation includes:
1. A bug fix in strbuf_reencode() that corrects the allocation size passed to strbuf_attach(), passing len+1 instead of len so that the existing buffer is reused rather than immediately reallocated.
2. A cleanup of strbuf_attach() call sites that were passing alloc == len, leaving no room for the NUL terminator. Sites with a known-NUL-terminated buffer now pass len+1; sites where the source buffer has no trailing NUL (ll_merge output) are converted to use strbuf_add() instead.
3. The main feature: HTTP 429 retry logic with support for the Retry-After header (both delay-seconds and HTTP-date formats), configurable via http.maxRetries, http.retryAfter, and http.maxRetryTime options. If any computed delay exceeds maxRetryTime the request fails immediately with a clear diagnostic rather than capping and retrying silently.
Vaidas Pilkauskas (3):
strbuf: pass correct alloc to strbuf_attach() in strbuf_reencode()
strbuf_attach: fix call sites to pass correct alloc
http: add support for HTTP 429 rate limit retries
Documentation/config/http.adoc | 26 ++++ builtin/am.c | 2 +- builtin/fast-import.c | 2 +- git-curl-compat.h | 8 + http.c | 144 +++++++++++++++--- http.h | 9 ++ mailinfo.c | 2 +- refs/files-backend.c | 2 +- remote-curl.c | 11 ++ strbuf.c | 2 +- t/lib-httpd.sh | 1 + t/lib-httpd/apache.conf | 8 + t/lib-httpd/http-429.sh | 98 ++++++++++++ t/meson.build | 1 + t/t5584-http-429-retry.sh | 266 +++++++++++++++++++++++++++++++++ trailer.c | 2 +- 16 files changed, 557 insertions(+), 27 deletions(-) create mode 100644 t/lib-httpd/http-429.sh create mode 100755 t/t5584-http-429-retry.sh
base-commit: ca1db8a0f7dc0dbea892e99f5b37c5fe5861be71
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-2008%2Fvaidas-shopify%2Fretry-after-v6
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-2008/vaidas-shopify/retry-after-v6
Pull-Request: https://github.com/gitgitgadget/git/pull/2008
Range-diff vs v5:
1: 7ec2d66447 = 1: 6e76be1d85 strbuf: pass correct alloc to strbuf_attach() in strbuf_reencode()
2: 3e0b78cfb6 = 2: 0dc214d3c2 strbuf_attach: fix call sites to pass correct alloc
3: 973703e9dd < -: ---------- remote-curl: introduce show_http_message_fatal() helper
4: bfee1f10c0 ! 3: 3418f4553d http: add support for HTTP 429 rate limit retries
@@ Documentation/config/http.adoc: http.keepAliveCount::
+ Default wait time in seconds before retrying when a server returns
+ HTTP 429 (Too Many Requests) without a Retry-After header.
+ Defaults to 0 (retry immediately). When a Retry-After header is
-+ present, its value takes precedence over this setting. Can be
-+ overridden by the `GIT_HTTP_RETRY_AFTER` environment variable.
++ present, its value takes precedence over this setting; however,
++ automatic use of the server-provided `Retry-After` header requires
++ libcurl 7.66.0 or later. On older versions, configure this setting
++ manually to control the retry delay. Can be overridden by the
++ `GIT_HTTP_RETRY_AFTER` environment variable.
+ See also `http.maxRetries` and `http.maxRetryTime`.
+
+http.maxRetries::
@@ http.c: static inline int is_hdr_continuation(const char *ptr, const size_t size
}
-static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UNUSED)
-+static size_t fwrite_headers(char *ptr, size_t eltsize, size_t nmemb, void *p MAYBE_UNUSED)
++static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p MAYBE_UNUSED)
{
size_t size = eltsize * nmemb;
struct strvec *values = &http_auth.wwwauth_headers;
-@@ http.c: static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UN
- goto exit;
- }
-
-+#ifndef GIT_CURL_HAVE_CURLINFO_RETRY_AFTER
-+ /* Parse Retry-After header for rate limiting (for curl < 7.66.0) */
-+ if (skip_iprefix_mem(ptr, size, "retry-after:", &val, &val_len)) {
-+ struct active_request_slot *slot = (struct active_request_slot *)p;
-+
-+ strbuf_add(&buf, val, val_len);
-+ strbuf_trim(&buf);
-+
-+ if (slot && slot->results) {
-+ /* Parse the retry-after value (delay-seconds or HTTP-date) */
-+ char *endptr;
-+ long retry_after;
-+
-+ errno = 0;
-+ retry_after = strtol(buf.buf, &endptr, 10);
-+
-+ /* Check if it's a valid integer (delay-seconds format) */
-+ if (endptr != buf.buf && *endptr == '\0' &&
-+ errno != ERANGE && retry_after >= 0) {
-+ slot->results->retry_after = retry_after;
-+ } else {
-+ /* Try parsing as HTTP-date format */
-+ timestamp_t timestamp;
-+ int offset;
-+ if (!parse_date_basic(buf.buf, ×tamp, &offset)) {
-+ /* Successfully parsed as date, calculate delay from now */
-+ timestamp_t now = time(NULL);
-+ if (timestamp > now) {
-+ slot->results->retry_after = (long)(timestamp - now);
-+ } else {
-+ /* Past date means retry immediately */
-+ slot->results->retry_after = 0;
-+ }
-+ } else {
-+ /* Failed to parse as either delay-seconds or HTTP-date */
-+ warning(_("unable to parse Retry-After header value: '%s'"), buf.buf);
-+ }
-+ }
-+ }
-+
-+ goto exit;
-+ }
-+#endif
-+
- /*
- * This line could be a continuation of the previously matched header
- * field. If this is the case then we should append this value to the
-@@ http.c: static void finish_active_slot(struct active_request_slot *slot)
-
- curl_easy_getinfo(slot->curl, CURLINFO_HTTP_CONNECTCODE,
- &slot->results->http_connectcode);
-+
-+#ifdef GIT_CURL_HAVE_CURLINFO_RETRY_AFTER
-+ if (slot->results->http_code == 429) {
-+ curl_off_t retry_after;
-+ CURLcode res = curl_easy_getinfo(slot->curl,
-+ CURLINFO_RETRY_AFTER,
-+ &retry_after);
-+ if (res == CURLE_OK && retry_after > 0)
-+ slot->results->retry_after = (long)retry_after;
-+ }
-+#endif
- }
-
- /* Run callback if appropriate */
@@ http.c: static int http_options(const char *var, const char *value,
return 0;
}
@@ http.c: int run_one_slot(struct active_request_slot *slot,
struct slot_results *results)
{
slot->results = results;
-+ /* Initialize retry_after to -1 (not set) */
-+ results->retry_after = -1;
+
if (!start_active_slot(slot)) {
xsnprintf(curl_errorstr, sizeof(curl_errorstr),
@@ http.c: static void http_opt_request_remainder(CURL *curl, off_t pos)
static int http_request(const char *url,
void *result, int target,
- const struct http_get_options *options)
-+ const struct http_get_options *options,
-+ long *retry_after_out)
++ struct http_get_options *options)
{
struct active_request_slot *slot;
- struct slot_results results;
+- struct slot_results results;
++ struct slot_results results = { .retry_after = -1 };
+ struct curl_slist *headers = http_copy_default_headers();
+ struct strbuf buf = STRBUF_INIT;
+ const char *accept_language;
@@ http.c: static int http_request(const char *url,
- fwrite_buffer);
- }
+ headers = curl_slist_append(headers, accept_language);
-- curl_easy_setopt(slot->curl, CURLOPT_HEADERFUNCTION, fwrite_wwwauth);
-+ curl_easy_setopt(slot->curl, CURLOPT_HEADERFUNCTION, fwrite_headers);
-+ curl_easy_setopt(slot->curl, CURLOPT_HEADERDATA, slot);
+ strbuf_addstr(&buf, "Pragma:");
+- if (options && options->no_cache)
++ if (options->no_cache)
+ strbuf_addstr(&buf, " no-cache");
+- if (options && options->initial_request &&
++ if (options->initial_request &&
+ http_follow_config == HTTP_FOLLOW_INITIAL)
+ curl_easy_setopt(slot->curl, CURLOPT_FOLLOWLOCATION, 1L);
- accept_language = http_get_accept_language_header();
+ headers = curl_slist_append(headers, buf.buf);
+ /* Add additional headers here */
+- if (options && options->extra_headers) {
++ if (options->extra_headers) {
+ const struct string_list_item *item;
+- if (options && options->extra_headers) {
+- for_each_string_list_item(item, options->extra_headers) {
+- headers = curl_slist_append(headers, item->string);
+- }
+- }
++ for_each_string_list_item(item, options->extra_headers)
++ headers = curl_slist_append(headers, item->string);
+ }
+
+ headers = http_append_auth_header(&http_auth, headers);
@@ http.c: static int http_request(const char *url,
ret = run_one_slot(slot, &results);
-+ /* Store retry_after from slot results if output parameter provided */
-+ if (retry_after_out)
-+ *retry_after_out = results.retry_after;
+- if (options && options->content_type) {
++#ifdef GIT_CURL_HAVE_CURLINFO_RETRY_AFTER
++ if (ret == HTTP_RATE_LIMITED) {
++ curl_off_t retry_after;
++ if (curl_easy_getinfo(slot->curl, CURLINFO_RETRY_AFTER,
++ &retry_after) == CURLE_OK && retry_after > 0)
++ results.retry_after = (long)retry_after;
++ }
++#endif
++
++ options->retry_after = results.retry_after;
+
- if (options && options->content_type) {
++ if (options->content_type) {
struct strbuf raw = STRBUF_INIT;
curlinfo_strbuf(slot->curl, CURLINFO_CONTENT_TYPE, &raw);
+ extract_content_type(&raw, options->content_type,
+@@ http.c: static int http_request(const char *url,
+ strbuf_release(&raw);
+ }
+
+- if (options && options->effective_url)
++ if (options->effective_url)
+ curlinfo_strbuf(slot->curl, CURLINFO_EFFECTIVE_URL,
+ options->effective_url);
+
@@ http.c: static int update_url_from_redirect(struct strbuf *base,
return 1;
}
-static int http_request_reauth(const char *url,
+/*
-+ * Handle rate limiting retry logic for HTTP 429 responses.
-+ * Returns a negative value if retries are exhausted or configuration is invalid,
-+ * otherwise returns the delay value (>= 0) to indicate the retry should proceed.
++ * Compute the retry delay for an HTTP 429 response.
++ * Returns a negative value if configuration is invalid (delay exceeds
++ * http.maxRetryTime), otherwise returns the delay in seconds (>= 0).
+ */
-+static long handle_rate_limit_retry(int *rate_limit_retries, long slot_retry_after)
++static long handle_rate_limit_retry(long slot_retry_after)
+{
-+ int retry_attempt = http_max_retries - *rate_limit_retries + 1;
-+
-+ trace2_data_intmax("http", the_repository, "http/429-retry-attempt",
-+ retry_attempt);
-+
-+ if (*rate_limit_retries <= 0) {
-+ /* Retries are disabled or exhausted */
-+ if (http_max_retries > 0) {
-+ error(_("too many rate limit retries, giving up"));
-+ trace2_data_string("http", the_repository,
-+ "http/429-error", "retries-exhausted");
-+ }
-+ return -1;
-+ }
-+
-+ (*rate_limit_retries)--;
-+
+ /* Use the slot-specific retry_after value or configured default */
+ if (slot_retry_after >= 0) {
+ /* Check if retry delay exceeds maximum allowed */
@@ http.c: static int update_url_from_redirect(struct strbuf *base,
void *result, int target,
struct http_get_options *options)
{
++ static struct http_get_options empty_opts;
int i = 3;
int ret;
+ int rate_limit_retries = http_max_retries;
-+ long slot_retry_after = -1; /* Per-slot retry_after value */
++
++ if (!options)
++ options = &empty_opts;
if (always_auth_proactively())
credential_fill(the_repository, &http_auth, 1);
-- ret = http_request(url, result, target, options);
-+ ret = http_request(url, result, target, options, &slot_retry_after);
+ ret = http_request(url, result, target, options);
- if (ret != HTTP_OK && ret != HTTP_REAUTH)
+ if (ret != HTTP_OK && ret != HTTP_REAUTH && ret != HTTP_RATE_LIMITED)
return ret;
+- if (options && options->effective_url && options->base_url) {
+ /* If retries are disabled and we got a 429, fail immediately */
+ if (ret == HTTP_RATE_LIMITED && !http_max_retries)
+ return HTTP_ERROR;
+
- if (options && options->effective_url && options->base_url) {
++ if (options->effective_url && options->base_url) {
if (update_url_from_redirect(options->base_url,
url, options->effective_url)) {
+ credential_from_url(&http_auth, options->base_url->buf);
@@ http.c: static int http_request_reauth(const char *url,
}
}
- while (ret == HTTP_REAUTH && --i) {
-+ while ((ret == HTTP_REAUTH || ret == HTTP_RATE_LIMITED) && --i) {
++ while ((ret == HTTP_REAUTH && --i) ||
++ (ret == HTTP_RATE_LIMITED && --rate_limit_retries)) {
+ long retry_delay = -1;
/*
* The previous request may have put cruft into our output stream; we
@@ http.c: static int http_request_reauth(const char *url,
default:
BUG("Unknown http_request target");
}
+-
+- credential_fill(the_repository, &http_auth, 1);
+ if (ret == HTTP_RATE_LIMITED) {
-+ retry_delay = handle_rate_limit_retry(&rate_limit_retries, slot_retry_after);
++ retry_delay = handle_rate_limit_retry(options->retry_after);
+ if (retry_delay < 0)
+ return HTTP_ERROR;
+
@@ http.c: static int http_request_reauth(const char *url,
+ "http/retry-sleep-seconds", retry_delay);
+ sleep(retry_delay);
+ }
-+ slot_retry_after = -1; /* Reset after use */
+ } else if (ret == HTTP_REAUTH) {
+ credential_fill(the_repository, &http_auth, 1);
+ }
-- credential_fill(the_repository, &http_auth, 1);
--
-- ret = http_request(url, result, target, options);
-+ ret = http_request(url, result, target, options, &slot_retry_after);
+ ret = http_request(url, result, target, options);
}
++ if (ret == HTTP_RATE_LIMITED) {
++ trace2_data_string("http", the_repository,
++ "http/429-error", "retries-exhausted");
++ return HTTP_RATE_LIMITED;
++ }
return ret;
}
+
@@ http.c: int http_get_strbuf(const char *url,
struct strbuf *result,
struct http_get_options *options)
@@ http.h: struct slot_results {
};
struct active_request_slot {
+@@ http.h: struct http_get_options {
+ * request has completed.
+ */
+ struct string_list *extra_headers;
++
++ /*
++ * After a request completes, contains the Retry-After delay in seconds
++ * if the server returned HTTP 429 with a Retry-After header (requires
++ * libcurl 7.66.0 or later), or -1 if no such header was present.
++ */
++ long retry_after;
+ };
+
+ /* Return values for http_get_*() */
@@ http.h: struct http_get_options {
#define HTTP_REAUTH 4
#define HTTP_NOAUTH 5
@@ http.h: struct http_get_options {
## remote-curl.c ##
@@ remote-curl.c: static struct discovery *discover_refs(const char *service, int for_push)
- show_http_message_fatal(&type, &charset, &buffer,
- _("unable to access '%s' with http.pinnedPubkey configuration: %s"),
- transport_anonymize_url(url.buf), curl_errorstr);
+ show_http_message(&type, &charset, &buffer);
+ die(_("unable to access '%s' with http.pinnedPubkey configuration: %s"),
+ transport_anonymize_url(url.buf), curl_errorstr);
+ case HTTP_RATE_LIMITED:
-+ show_http_message_fatal(&type, &charset, &buffer,
-+ _("rate limited by '%s', please try again later"),
-+ transport_anonymize_url(url.buf));
++ if (http_options.retry_after > 0) {
++ show_http_message(&type, &charset, &buffer);
++ die(_("rate limited by '%s', please try again in %ld seconds"),
++ transport_anonymize_url(url.buf),
++ http_options.retry_after);
++ } else {
++ show_http_message(&type, &charset, &buffer);
++ die(_("rate limited by '%s', please try again later"),
++ transport_anonymize_url(url.buf));
++ }
default:
- show_http_message_fatal(&type, &charset, &buffer,
- _("unable to access '%s': %s"),
+ show_http_message(&type, &charset, &buffer);
+ die(_("unable to access '%s': %s"),
## t/lib-httpd.sh ##
@@ t/lib-httpd.sh: prepare_httpd() {
@@ t/t5584-http-429-retry.sh (new)
+ test_must_fail git ls-remote "$HTTPD_URL/http_429/retry-after-exceeds-max-time/100/repo.git" 2>err &&
+ duration=$(test-tool date getnanos $start) &&
+
-+ # Should fail quickly (less than 2 seconds, no 100 second wait)
++ # Should fail quickly (no 100 second wait)
+ duration_int=${duration%.*} &&
-+ test "$duration_int" -lt 2 &&
++ test "$duration_int" -lt 99 &&
+ test_grep "greater than http.maxRetryTime" err
+'
+
@@ t/t5584-http-429-retry.sh (new)
+ test_must_fail git ls-remote "$HTTPD_URL/http_429/config-retry-after-exceeds-max-time/none/repo.git" 2>err &&
+ duration=$(test-tool date getnanos $start) &&
+
-+ # Should fail quickly
++ # Should fail quickly (no 100 second wait)
+ duration_int=${duration%.*} &&
-+ test "$duration_int" -lt 2 &&
++ test "$duration_int" -lt 99 &&
+ test_grep "configured http.retryAfter.*exceeds.*http.maxRetryTime" err
+'
+
@@ t/t5584-http-429-retry.sh (new)
+
+ # Should fail quickly (not wait 200 seconds)
+ duration_int=${duration%.*} &&
-+ test "$duration_int" -lt 2 &&
++ test "$duration_int" -lt 199 &&
+ test_grep "http.maxRetryTime" err
+'
+
@@ t/t5584-http-429-retry.sh (new)
+ git ls-remote "$HTTPD_URL/http_429/past-http-date/$past_date_encoded/repo.git" >output 2>err &&
+ duration=$(test-tool date getnanos $start) &&
+
-+ # Should complete quickly (less than 2 seconds)
++ # Should complete quickly (no wait for a past-date Retry-After)
+ duration_int=${duration%.*} &&
-+ test "$duration_int" -lt 2 &&
++ test "$duration_int" -lt 5 &&
+ test_grep "refs/heads/" output
+'
+
@@ t/t5584-http-429-retry.sh (new)
+
+ # Should fail quickly (not wait 50 seconds) because env var limits to 10
+ duration_int=${duration%.*} &&
-+ test "$duration_int" -lt 5 &&
++ test "$duration_int" -lt 49 &&
+ test_grep "greater than http.maxRetryTime" err
+'
+Signature · 1 line
gitgitgadget
PATCH v6 2/3 strbuf_attach: fix call sites to pass correct alloc
From: Vaidas Pilkauskas <[email protected]>
strbuf_attach(sb, buf, len, alloc) requires alloc > len (the buffer must have at least len+1 bytes to hold the NUL). Several call sites passed alloc == len, relying on strbuf_grow(sb, 0) inside strbuf_attach to reallocate. Fix these in mailinfo, am, refs/files-backend, fast-import, and trailer by passing len+1 when the buffer is a NUL-terminated string (or from strbuf_detach).
--- builtin/am.c | 2 +- builtin/fast-import.c | 2 +- mailinfo.c | 2 +- refs/files-backend.c | 2 +- trailer.c | 2 +- 5 files changed, 5 insertions(+), 5 deletions(-)
builtin/am.c +1 -1
@@ -1188,7 +1188,7 @@ static void am_append_signoff(struct am_state *state) {struct strbuf sb =STRBUF_INIT ; -strbuf_attach (&sb, state->msg, state->msg_len, state->msg_len); +strbuf_attach (&sb, state->msg, state->msg_len, state->msg_len +1 );append_signoff (&sb,0 ,0 ); state->msg =strbuf_detach (&sb, &state->msg_len); }
builtin/fast-import.c +1 -1
@@ -3250,7 +3250,7 @@ static void cat_blob(struct object_entry *oe, struct object_id *oid)cat_blob_write ("\n" ,1 );if (oe && oe->pack_id == pack_id) { last_blob.offset = oe->idx.offset; -strbuf_attach (&last_blob.data, buf, size, size); +strbuf_attach (&last_blob.data, buf, size, size +1 ); last_blob.depth = oe->depth; }else free (buf);
mailinfo.c +1 -1
@@ -470,7 +470,7 @@ static int convert_to_utf8(struct mailinfo *mi,return error ("cannot convert from %s to %s" , charset, mi->metainfo_charset); } -strbuf_attach (line, out, out_len, out_len); +strbuf_attach (line, out, out_len, out_len +1 );return 0 ; }
refs/files-backend.c +1 -1
@@ -1813,7 +1813,7 @@ static int commit_ref(struct ref_lock *lock)size_t len =strlen (path);struct strbuf sb_path =STRBUF_INIT ; -strbuf_attach (&sb_path, path, len, len); +strbuf_attach (&sb_path, path, len, len +1 );/* * If this fails, commit_lock_file() will also fail
trailer.c +1 -1
@@ -1009,7 +1009,7 @@ static struct trailer_block *trailer_block_get(const struct process_trailer_optifor (ptr = trailer_lines; *ptr; ptr++) {if (last &&isspace ((*ptr)->buf[0 ])) {struct strbuf sb =STRBUF_INIT ; -strbuf_attach (&sb, *last,strlen (*last),strlen (*last)); +strbuf_attach (&sb, *last,strlen (*last),strlen (*last) +1 );strbuf_addbuf (&sb, *ptr); *last =strbuf_detach (&sb,NULL );continue ;
Signature · 1 line
gitgitgadget
PATCH v6 3/3 http: add support for HTTP 429 rate limit retries
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
Add retry logic for HTTP 429 (Too Many Requests) responses to handle server-side rate limiting gracefully. When Git's HTTP client receives a 429 response, it can now automatically retry the request after an appropriate delay, respecting the server's rate limits.
The implementation supports the RFC-compliant Retry-After header in both delay-seconds (integer) and HTTP-date (RFC 2822) formats. If a past date is provided, Git retries immediately without waiting.
Retry behavior is controlled by three new configuration options (http.maxRetries, http.retryAfter, and http.maxRetryTime) which are documented in git-config(1).
The retry logic implements a fail-fast approach: if any delay (whether from server header or configuration) exceeds maxRetryTime, Git fails immediately with a clear error message rather than capping the delay. This provides better visibility into rate limiting issues.
The implementation includes extensive test coverage for basic retry behavior, Retry-After header formats (integer and HTTP-date), configuration combinations, maxRetryTime limits, invalid header handling, environment variable overrides, and edge cases.
--- Documentation/config/http.adoc | 26 ++++ git-curl-compat.h | 8 + http.c | 144 +++++++++++++++--- http.h | 9 ++ remote-curl.c | 11 ++ t/lib-httpd.sh | 1 + t/lib-httpd/apache.conf | 8 + t/lib-httpd/http-429.sh | 98 ++++++++++++ t/meson.build | 1 + t/t5584-http-429-retry.sh | 266 +++++++++++++++++++++++++++++++++ 10 files changed, 551 insertions(+), 21 deletions(-) create mode 100644 t/lib-httpd/http-429.sh create mode 100755 t/t5584-http-429-retry.sh
Documentation/config/http.adoc +26 -0
@@ -315,6 +315,32 @@ http.keepAliveCount:: unset, curl's default value is used. Can be overridden by the `GIT_HTTP_KEEPALIVE_COUNT` environment variable. +http.retryAfter:: + Default wait time in seconds before retrying when a server returns + HTTP 429 (Too Many Requests) without a Retry-After header. + Defaults to 0 (retry immediately). When a Retry-After header is + present, its value takes precedence over this setting; however, + automatic use of the server-provided `Retry-After` header requires + libcurl 7.66.0 or later. On older versions, configure this setting + manually to control the retry delay. Can be overridden by the + `GIT_HTTP_RETRY_AFTER` environment variable. + See also `http.maxRetries` and `http.maxRetryTime`. + +http.maxRetries:: + Maximum number of times to retry after receiving HTTP 429 (Too Many + Requests) responses. Set to 0 (the default) to disable retries. + Can be overridden by the `GIT_HTTP_MAX_RETRIES` environment variable. + See also `http.retryAfter` and `http.maxRetryTime`. + +http.maxRetryTime:: + Maximum time in seconds to wait for a single retry attempt when + handling HTTP 429 (Too Many Requests) responses. If the server + requests a delay (via Retry-After header) or if `http.retryAfter` + is configured with a value that exceeds this maximum, Git will fail + immediately rather than waiting. Default is 300 seconds (5 minutes). + Can be overridden by the `GIT_HTTP_MAX_RETRY_TIME` environment + variable. See also `http.retryAfter` and `http.maxRetries`. + http.noEPSV:: A boolean which disables using of EPSV ftp command by curl. This can be helpful with some "poor" ftp servers which don't
git-curl-compat.h +8 -0
@@ -37,6 +37,14 @@#define GIT_CURL_NEED_TRANSFER_ENCODING_HEADER #endif +/** +* CURLINFO_RETRY_AFTER was added in 7.66.0, released in September 2019. +* It allows curl to automatically parse Retry-After headers. +*/ +#if LIBCURL_VERSION_NUM >=0x074200 +#define GIT_CURL_HAVE_CURLINFO_RETRY_AFTER 1 +#endif +/** * CURLOPT_PROTOCOLS_STR and CURLOPT_REDIR_PROTOCOLS_STR were added in 7.85.0, * released in August 2022.
http.c +123 -21
266 lines not on this page · raw message
http.h +9 -0
@@ -20,6 +20,7 @@ struct slot_results {long http_code;long auth_avail;long http_connectcode; +long retry_after; };struct active_request_slot {
@@ -157,6 +158,13 @@ struct http_get_options {* request has completed. */ struct string_list *extra_headers; + +/* +* After a request completes, contains the Retry-After delay in seconds +* if the server returned HTTP 429 with a Retry-After header (requires +* libcurl 7.66.0 or later), or -1 if no such header was present. +*/ +long retry_after; };/* Return values for http_get_*() */
@@ -167,6 +175,7 @@ struct http_get_options {#define HTTP_REAUTH 4#define HTTP_NOAUTH 5#define HTTP_NOMATCHPUBLICKEY 6 +#define HTTP_RATE_LIMITED 7/* * Requests a URL and stores the result in a strbuf.
remote-curl.c +11 -0
@@ -529,6 +529,17 @@ static struct discovery *discover_refs(const char *service, int for_push)show_http_message (&type, &charset, &buffer);die (_ ("unable to access '%s' with http.pinnedPubkey configuration: %s" ),transport_anonymize_url (url.buf), curl_errorstr); +case HTTP_RATE_LIMITED : +if (http_options.retry_after >0 ) { +show_http_message (&type, &charset, &buffer); +die (_ ("rate limited by '%s', please try again in %ld seconds" ), +transport_anonymize_url (url.buf), + http_options.retry_after); + }else { +show_http_message (&type, &charset, &buffer); +die (_ ("rate limited by '%s', please try again later" ), +transport_anonymize_url (url.buf)); + }default :show_http_message (&type, &charset, &buffer); die(_("unable to access '%s': %s"),
t/lib-httpd.sh +1 -0
@@ -167,6 +167,7 @@ prepare_httpd() { install_script error.sh install_script apply-one-time-script.sh install_script nph-custom-auth.sh + install_script http-429.sh ln -s "$LIB_HTTPD_MODULE_PATH" "$HTTPD_ROOT_PATH/modules"
t/lib-httpd/apache.conf +8 -0
@@ -139,6 +139,10 @@ SetEnv PERL_PATH ${PERL_PATH} SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} SetEnv GIT_HTTP_EXPORT_ALL </LocationMatch> +<LocationMatch /http_429/> + SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} + SetEnv GIT_HTTP_EXPORT_ALL +</LocationMatch> <LocationMatch /smart_v0/> SetEnv GIT_EXEC_PATH ${GIT_EXEC_PATH} SetEnv GIT_HTTP_EXPORT_ALL
@@ -160,6 +164,7 @@ ScriptAlias /broken_smart/ broken-smart-http.sh/ ScriptAlias /error_smart/ error-smart-http.sh/ ScriptAlias /error/ error.sh/ ScriptAliasMatch /one_time_script/(.*) apply-one-time-script.sh/$1 +ScriptAliasMatch /http_429/(.*) http-429.sh/$1 ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1 <Directory ${GIT_EXEC_PATH}> Options FollowSymlinks
@@ -185,6 +190,9 @@ ScriptAliasMatch /custom_auth/(.*) nph-custom-auth.sh/$1 <Files apply-one-time-script.sh> Options ExecCGI </Files> +<Files http-429.sh> + Options ExecCGI +</Files> <Files ${GIT_EXEC_PATH}/git-http-backend> Options ExecCGI </Files>
t/lib-httpd/http-429.sh new file +98 -0
98 lines not on this page · raw message
t/meson.build +1 -0
@@ -709,6 +709,7 @@ integration_tests = [ 't5581-http-curl-verbose.sh', 't5582-fetch-negative-refspec.sh', 't5583-push-branches.sh', + 't5584-http-429-retry.sh', 't5600-clone-fail-cleanup.sh', 't5601-clone.sh', 't5602-clone-remote-exec.sh',
t/t5584-http-429-retry.sh new file +266 -0
266 lines not on this page · raw message
Signature · 1 line
gitgitgadget
PATCH v6 1/3 strbuf: pass correct alloc to strbuf_attach() in strbuf_reencode()
In reply to Vaidas Pilkauskas via GitGitGadgetFrom: Vaidas Pilkauskas <[email protected]>
reencode_string_len() allocates len+1 bytes (including the NUL) and returns the string length in len. strbuf_reencode() was calling strbuf_attach(sb, out, len, len), so alloc was one byte too small.
strbuf_attach() then calls strbuf_grow(sb, 0). With alloc < len+1, ALLOC_GROW always reallocates, so we reallocated immediately after attach even when the strbuf was not extended further. Pass len+1 as the alloc argument so the existing buffer is reused and the reallocation is avoided.
--- strbuf.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
strbuf.c +1 -1
@@ -168,7 +168,7 @@ int strbuf_reencode(struct strbuf *sb, const char *from, const char *to)if (!out)return -1 ; -strbuf_attach (sb, out, len, len); +strbuf_attach (sb, out, len, len +1 );return 0 ; }
Signature · 1 line
gitgitgadget
http.c static int http_options(const char *var, const char *value,
Quoted 9 lines · Vaidas Pilkauskas via GitGitGadget
http.c +1
-1return size && (*ptr ==' ' || *ptr =='\t' ); } -static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p UNUSED) +static size_t fwrite_wwwauth(char *ptr, size_t eltsize, size_t nmemb, void *p MAYBE_UNUSED) {size_t size = eltsize * nmemb;struct strvec *values = &http_auth.wwwauth_headers;@@ -575,6 +582,21 @@ static int http_options(const char *var, const char *value,
Good, this version drops the special case where we do not define GIT_CURL_HAVE_CURLINFO_RETRY_AFTER, which Peff suggested in his review of the earlier round.
I agree with his suggestion that we can document that handling Retry-After requires a libcurl newer than 7.66.0, and that is well documented in the user-facing documentation and code comments where appropriate.
http.c static void http_opt_request_remainder(CURL *curl, off_t pos) · line 2153
Quoted 6 lines
http.c +1
-1@@ -2119,10 +2150,10 @@ static void http_opt_request_remainder(CURL *curl, off_t pos)static int http_request (const char *url,void *result,int target, -const struct http_get_options *options) +struct http_get_options *options)
The previous round had this as a const pointer, with a separate out-parameter via 'long *retry_after_out'. Review on the previous round suggested making the retry_after part of the existing out-parameter. Of course, doing so requires that we make that parameter non-const, hence the change here, which looks good to me.
http.c static void http_opt_request_remainder(CURL *curl, off_t pos) · line 2156
Quoted 4 lines
http.c +1
-1{struct active_request_slot *slot; -struct slot_results results; +struct slot_results results = { .retry_after =-1 };
This also moved from run_one_slot(); this location makes more sense to me.
Quoted 13 lines
http.h +1
-0@@ -20,6 +20,7 @@ struct slot_results {long http_code;long auth_avail;long http_connectcode; +long retry_after; };struct active_request_slot {
http.h struct http_get_options { · line 167
Quoted 10 lines
http.h +7
-0@@ -157,6 +158,13 @@ struct http_get_options {* request has completed. */ struct string_list *extra_headers; + +/* +* After a request completes, contains the Retry-After delay in seconds +* if the server returned HTTP 429 with a Retry-After header (requires +* libcurl 7.66.0 or later), or -1 if no such header was present. +*/ +long retry_after;
I think making this a pure long instead of a pointer as is the case with other members of this struct makes sense for the reasons that Peff pointed out in the review of the previous round.
Thanks,
Taylor
Quoted 9 lines · Vaidas Pilkauskas via GitGitGadget
Changes since v5:
* drop
show_http_message_fatal()preparation patch
* dropfwrite_headers, restorefwrite_wwwauth
* moveCURLINFO_RETRY_AFTERfromfinish_active_slottohttp_request
* move retry_after=-1 init fromrun_one_slottohttp_request
* replaceretry_after_outparam withhttp_get_optionsfield
* fix loop counter: separate REAUTH andRATE_LIMITEDcounters
* fix racy -lt 2 timing bounds in tests
Thanks, this round looks good to me. The main things that I noted from the review on v5 was to drop the old 3/4, and a handful of suggestions on the final patch, all of which look to have been addressed.
The first two patches being unchanged, this round looks good to me. Thanks for working on this, Vaidas!
Thanks,
Taylor
Quoted 15 lines · Taylor Blau
Quoted 9 lines · Vaidas Pilkauskas via GitGitGadget
Changes since v5:
* drop
show_http_message_fatal()preparation patch
* dropfwrite_headers, restorefwrite_wwwauth
* moveCURLINFO_RETRY_AFTERfromfinish_active_slottohttp_request
* move retry_after=-1 init fromrun_one_slottohttp_request
* replaceretry_after_outparam withhttp_get_optionsfield
* fix loop counter: separate REAUTH andRATE_LIMITEDcounters
* fix racy -lt 2 timing bounds in testsThanks, this round looks good to me. The main things that I noted from the review on v5 was to drop the old 3/4, and a handful of suggestions on the final patch, all of which look to have been addressed.
Quoted 5 lines
The first two patches being unchanged, this round looks good to me. Thanks for working on this, Vaidas!
Thanks,
Taylor
Thanks, both.
Let me mark the topic for 'next' then.
Quoted 10 lines · Taylor Blau
Thanks, this round looks good to me. The main things that I noted from the review on v5 was to drop the old 3/4, and a handful of suggestions on the final patch, all of which look to have been addressed.
The first two patches being unchanged, this round looks good to me. Thanks for working on this, Vaidas!
Thanks,
Taylor
Taylor, Peff and Junio, thank you very much for reviewing!