Noticed this recently when trying to run a sharness test while under a <code class="no

Ah, this is likely the problem: <div class="Box Box--condensed m

`flux start` hangs when run under `flux proxy` about flux-core HOT 2 CLOSED

grondo commented on September 9, 2024

`flux start` hangs when run under `flux proxy`

from flux-core.

Comments (2)

grondo commented on September 9, 2024

Ah, this is likely the problem:

flux-core/src/common/libflux/attr.c

Lines 107 to 149 in 6fe8a9b

 if (!(c = get_attr_cache (h))) 

 return NULL; 

 /* The attribute parent-uri is treated specifically here, since 

  * a process connected to this instance via flux-proxy(1) should 

  * have the parent-uri returned as a usable remote URI. 

  * 

  * Therefore, if FLUX_PROXY_REMOTE is set in the current environment, 

  * and the attribute is parent-uri, instead actually return the 

  * handle-specific parent-remote-uri attribute. If this is not yet 

  * available, fetch parent-uri and construct the remote uri by 

  * substituting FLUX_PROXY_REMOTE in an ssh:// uri. 

  */ 

 if (streq (name, "parent-uri") 

 && (proxy_remote = getenv ("FLUX_PROXY_REMOTE"))) { 

 orig_name = name; 

 name = "parent-remote-uri"; 

 } 

 if ((val = zhashx_lookup (c->cache, name))) 

 return val; 

 if (!(f = flux_rpc_pack (h, 

 "attr.get", 

 FLUX_NODEID_ANY, 

 0, 

 "{s:s}", 

 "name", orig_name))) 

 return NULL; 

 if (flux_rpc_get_unpack (f, 

 "{s:s s:i}", 

 "value", &val, 

 "flags", &flags) < 0) 

 goto done; 

 /* If proxy_remote is non-NULL then parent-uri has been aliased to 

  * parent-remote-uri. Swap a local URI to a remote: 

  */ 

 if (proxy_remote 

 && strstarts (val, "local://")) { 

 if (asprintf (&cpy, "ssh://%s%s", proxy_remote, val+8) < 0) 

 goto done; 

 } 

 else if (!(cpy = strdup (val))) 

 goto done;

when the parent-uri attribute is fetched, flux_attr_get(3) checks for an environment variable set by flux-proxy and attempts to instead provide a remote uri. However, the name parent-uri is updated to parent-remote-uri before checking the cache, so if parent-remote-uri is not already cached, then this call always results in an RPC, even if parent-uri is cached.

This breaks the assumption in broker/module.c attr_cache_to_json() that all attrs returned by flux_attr_cache_first() and flux_attr_cache_next() will not result in an RPC.

Additionally, even if we fixed the cache check to avoid the RPC, the wrong value for parent-uri would be set in the json representation of the cache passed to the module since it will have been translated to a remote URI.

Probably the right solution for now is to just unset FLUX_PROXY_REMOTE in the broker (along with FLUX_URI), since once another broker is started, the broker and its descendants are no longer under the influence of a remote flux proxy. However, probably the cache check noted above should also be fixed, and maybe the whole way the uri is auto-translated to its remote equivalent should have another think...

from flux-core.

grondo commented on September 9, 2024

I did verify that unsetenv ("FLUX_PROXY_REMOTE"); early in the broker avoids this issue.

from flux-core.

`flux start` hangs when run under `flux proxy` about flux-core HOT 2 CLOSED

Comments (2)

Related Issues (20)

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

	if (!(c = get_attr_cache (h)))
	return NULL;

	/* The attribute parent-uri is treated specifically here, since
	* a process connected to this instance via flux-proxy(1) should
	* have the parent-uri returned as a usable remote URI.
	*
	* Therefore, if FLUX_PROXY_REMOTE is set in the current environment,
	* and the attribute is parent-uri, instead actually return the
	* handle-specific parent-remote-uri attribute. If this is not yet
	* available, fetch parent-uri and construct the remote uri by
	* substituting FLUX_PROXY_REMOTE in an ssh:// uri.
	*/
	if (streq (name, "parent-uri")
	&& (proxy_remote = getenv ("FLUX_PROXY_REMOTE"))) {
	orig_name = name;
	name = "parent-remote-uri";
	}
	if ((val = zhashx_lookup (c->cache, name)))
	return val;
	if (!(f = flux_rpc_pack (h,
	"attr.get",
	FLUX_NODEID_ANY,
	0,
	"{s:s}",
	"name", orig_name)))
	return NULL;
	if (flux_rpc_get_unpack (f,
	"{s:s s:i}",
	"value", &val,
	"flags", &flags) < 0)
	goto done;

	/* If proxy_remote is non-NULL then parent-uri has been aliased to
	* parent-remote-uri. Swap a local URI to a remote:
	*/
	if (proxy_remote
	&& strstarts (val, "local://")) {
	if (asprintf (&cpy, "ssh://%s%s", proxy_remote, val+8) < 0)
	goto done;
	}
	else if (!(cpy = strdup (val)))
	goto done;