[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[bug#66573] [PATCH v2 06/11] gnu: ucx: Update to 1.15.0.
From: |
Ludovic Courtès |
Subject: |
[bug#66573] [PATCH v2 06/11] gnu: ucx: Update to 1.15.0. |
Date: |
Thu, 19 Oct 2023 21:36:16 +0200 |
* gnu/packages/fabric-management.scm (ucx): Update to 1.15.0.
* gnu/packages/patches/ucx-tcp-iface-ioctl.patch: Update for 1.15.0.
---
gnu/packages/fabric-management.scm | 4 +-
.../patches/ucx-tcp-iface-ioctl.patch | 105 +++++++++++-------
2 files changed, 64 insertions(+), 45 deletions(-)
diff --git a/gnu/packages/fabric-management.scm
b/gnu/packages/fabric-management.scm
index ccdaa0ee0a..f41b4e99ed 100644
--- a/gnu/packages/fabric-management.scm
+++ b/gnu/packages/fabric-management.scm
@@ -185,7 +185,7 @@ (define-public ibutils
(define-public ucx
(package
(name "ucx")
- (version "1.14.0")
+ (version "1.15.0")
(source (origin
(method git-fetch)
(uri (git-reference
@@ -195,7 +195,7 @@ (define-public ucx
(patches (search-patches "ucx-tcp-iface-ioctl.patch"))
(sha256
(base32
- "0ki2r768wqm92qv06wxrh3kv2nl2yj4ds9fz0s0b5rr2ycjiw9ir"))))
+ "1mk46vyfp8hsivk88s8gv0nf458jfs59fczpf66wwa3a9yp324jp"))))
(build-system gnu-build-system)
(arguments
(list
diff --git a/gnu/packages/patches/ucx-tcp-iface-ioctl.patch
b/gnu/packages/patches/ucx-tcp-iface-ioctl.patch
index c441a0861a..2a0e4ce138 100644
--- a/gnu/packages/patches/ucx-tcp-iface-ioctl.patch
+++ b/gnu/packages/patches/ucx-tcp-iface-ioctl.patch
@@ -3,102 +3,121 @@ TCP network interfaces cannot be obtained via
/sys/class/net. This patch
provides alternative code that uses the SIOCGIFCONF ioctl to get the
names of the available TCP network interfaces.
+Initially submitted at <https://github.com/openucx/ucx/pull/4462>.
+
diff --git a/src/uct/tcp/tcp_iface.c b/src/uct/tcp/tcp_iface.c
-index cad4a2709..7c1d2c9de 100644
+index 6a6cd34fa..af32bb2e9 100644
--- a/src/uct/tcp/tcp_iface.c
+++ b/src/uct/tcp/tcp_iface.c
-@@ -17,6 +17,8 @@
- #include <sys/poll.h>
+@@ -18,6 +18,8 @@
#include <netinet/tcp.h>
#include <dirent.h>
+ #include <float.h>
+#include <net/if.h>
+#include <sys/ioctl.h>
+ #define UCT_TCP_IFACE_NETDEV_DIR "/sys/class/net"
- extern ucs_class_t UCS_CLASS_DECL_NAME(uct_tcp_iface_t);
-@@ -586,6 +588,68 @@ static UCS_CLASS_DEFINE_NEW_FUNC(uct_tcp_iface_t,
uct_iface_t, uct_md_h,
+@@ -875,6 +877,85 @@ static UCS_CLASS_DEFINE_NEW_FUNC(uct_tcp_iface_t,
uct_iface_t, uct_md_h,
uct_worker_h, const uct_iface_params_t*,
const uct_iface_config_t*);
+/* Fetch information about available network devices through an ioctl. */
-+static ucs_status_t query_devices_ioctl(uct_md_h md,
-+ uct_tl_device_resource_t **tl_devices_p,
-+ unsigned *num_tl_devices_p)
++static ucs_status_t uct_tcp_query_devices_ioctl(uct_md_h md,
++ uct_tl_device_resource_t
**devices_p,
++ unsigned *num_devices_p)
+{
+ int sock, err, i;
-+ uct_tl_device_resource_t *resources, *tmp;
-+ unsigned num_resources;
++ uct_tl_device_resource_t *devices, *tmp;
++ unsigned num_devices;
+ ucs_status_t status;
+ struct ifconf conf;
-+ struct ifreq reqs[10];
+
-+ conf.ifc_len = sizeof reqs;
-+ conf.ifc_req = reqs;
++ conf.ifc_len = 0;
++ conf.ifc_req = NULL;
+
-+ sock = socket(SOCK_STREAM, AF_INET, 0);
-+ if (sock < 0) {
-+ ucs_error("socket(2) failed: %m");
-+ status = UCS_ERR_IO_ERROR;
-+ goto out;
++ status = ucs_socket_create(AF_INET, SOCK_STREAM, &sock);
++ if (status != UCS_OK) {
++ goto out;
+ }
+
+ err = ioctl(sock, SIOCGIFCONF, &conf);
+ if (err < 0) {
-+ ucs_error("SIOCGIFCONF ioctl failed: %m");
-+ status = UCS_ERR_IO_ERROR;
-+ goto out;
++ ucs_error("ioctl(SIOCGIFCONF) failed: %m");
++ status = UCS_ERR_IO_ERROR;
++ goto out;
+ }
+
-+ resources = NULL;
-+ num_resources = 0;
-+ for (i = 0; i < conf.ifc_len / sizeof(struct ifreq); i++) {
-+ const char *name = reqs[i].ifr_name;
++ conf.ifc_req = ucs_calloc(1, conf.ifc_len, "ifreq");
++ if (conf.ifc_req == NULL) {
++ ucs_error("memory alocation failed");
++ status = UCS_ERR_NO_MEMORY;
++ goto out;
++ }
++
++ err = ioctl(sock, SIOCGIFCONF, &conf);
++ if (err < 0) {
++ ucs_error("ioctl(SIOCGIFCONF) failed: %m");
++ status = UCS_ERR_IO_ERROR;
++ goto out_free;
++ }
++
++ devices = NULL;
++ num_devices = 0;
++ for (i = 0; i < (conf.ifc_len / sizeof(struct ifreq)); i++) {
++ const char *name = conf.ifc_req[i].ifr_name;
++ sa_family_t family = conf.ifc_req[i].ifr_addr.sa_family;
+
-+ if (!ucs_netif_is_active(name, AF_INET)) {
++ if (!ucs_netif_is_active(name, family)) {
+ continue;
+ }
+
-+ tmp = ucs_realloc(resources, sizeof(*resources) * (num_resources + 1),
-+ "tcp resources");
++ tmp = ucs_realloc(devices, sizeof(*devices) * (num_devices + 1),
++ "tcp devices");
+ if (tmp == NULL) {
-+ ucs_free(resources);
++ ucs_free(devices);
+ status = UCS_ERR_NO_MEMORY;
-+ goto out;
++ goto out_free;
+ }
-+ resources = tmp;
++ devices = tmp;
+
-+ ucs_snprintf_zero(resources[i].name, sizeof(resources[i].name),
++ ucs_snprintf_zero(devices[num_devices].name,
++ sizeof(devices[num_devices].name),
+ "%s", name);
-+ resources[i].type = UCT_DEVICE_TYPE_NET;
-+ ++num_resources;
++ devices[num_devices].type = UCT_DEVICE_TYPE_NET;
++ ++num_devices;
+ }
+
-+ *num_tl_devices_p = num_resources;
-+ *tl_devices_p = resources;
-+ status = UCS_OK;
++ *num_devices_p = num_devices;
++ *devices_p = devices;
++ status = UCS_OK;
+
++out_free:
++ ucs_free(conf.ifc_req);
+out:
-+ if (sock >= 0) close(sock);
++ if (sock >= 0) {
++ close(sock);
++ }
+ return status;
+}
+
ucs_status_t uct_tcp_query_devices(uct_md_h md,
uct_tl_device_resource_t **devices_p,
unsigned *num_devices_p)
-@@ -599,9 +663,9 @@ ucs_status_t uct_tcp_query_devices(uct_md_h md,
+@@ -893,9 +974,9 @@ ucs_status_t uct_tcp_query_devices(uct_md_h md,
dir = opendir(UCT_TCP_IFACE_NETDEV_DIR);
if (dir == NULL) {
- ucs_error("opendir(%s) failed: %m", UCT_TCP_IFACE_NETDEV_DIR);
- status = UCS_ERR_IO_ERROR;
- goto out;
-+ /* When /sys is unavailable, as can be the case in a container,
-+ * resort to a good old 'ioctl'. */
-+ return query_devices_ioctl(md, devices_p, num_devices_p);
++ /* When /sys is unavailable, as can be the case in a container,
++ * resort to a good old 'ioctl'. */
++ return uct_tcp_query_devices_ioctl(md, devices_p, num_devices_p);
}
devices = NULL;
-@@ -655,7 +719,6 @@ ucs_status_t uct_tcp_query_devices(uct_md_h md,
+@@ -963,7 +1044,6 @@ ucs_status_t uct_tcp_query_devices(uct_md_h md,
out_closedir:
closedir(dir);
--
2.41.0
- [bug#66573] [PATCH 07/11] gnu: rdma-core: Update to 48.0., (continued)
- [bug#66573] [PATCH 07/11] gnu: rdma-core: Update to 48.0., Ludovic Courtès, 2023/10/16
- [bug#66573] [PATCH 02/11] gnu: opensm: Use gexps., Ludovic Courtès, 2023/10/16
- [bug#66573] [PATCH 03/11] gnu: slurm: Add 23.02.6., Ludovic Courtès, 2023/10/16
- [bug#66573] [PATCH 05/11] gnu: opensm: Update to 3.3.24., Ludovic Courtès, 2023/10/16
- [bug#66573] [PATCH 01/11] gnu: opensm: Use the right version string for the doc directory., Ludovic Courtès, 2023/10/16
- [bug#66573] [PATCH 09/11] gnu: psm2: Update to 12.0., Ludovic Courtès, 2023/10/16
- [bug#66573] [PATCH 10/11] gnu: openmpi: Update to 4.1.6., Ludovic Courtès, 2023/10/16
- [bug#66573] [PATCH 11/11] gnu: intel-mpi-benchmarks: Update to 2021.3., Ludovic Courtès, 2023/10/16
- [bug#66573] [PATCH v2 02/11] gnu: opensm: Use gexps., Ludovic Courtès, 2023/10/19
- [bug#66573] [PATCH v2 01/11] gnu: opensm: Use the right version string for the doc directory., Ludovic Courtès, 2023/10/19
- [bug#66573] [PATCH v2 06/11] gnu: ucx: Update to 1.15.0.,
Ludovic Courtès <=
- [bug#66573] [PATCH v2 07/11] gnu: rdma-core: Update to 48.0., Ludovic Courtès, 2023/10/19
- [bug#66573] [PATCH v2 11/11] gnu: intel-mpi-benchmarks: Update to 2021.3., Ludovic Courtès, 2023/10/19
- [bug#66573] [PATCH v2 10/11] gnu: openmpi: Update to 4.1.6., Ludovic Courtès, 2023/10/19
- [bug#66573] [PATCH v2 04/11] gnu: hwloc: Update to 2.9.3., Ludovic Courtès, 2023/10/19
- [bug#66573] [PATCH v2 08/11] gnu: libfabric: Update to 1.19.0., Ludovic Courtès, 2023/10/19
- [bug#66573] [PATCH v2 03/11] gnu: slurm: Add 23.02.6., Ludovic Courtès, 2023/10/19
- [bug#66573] [PATCH v2 05/11] gnu: opensm: Update to 3.3.24., Ludovic Courtès, 2023/10/19
- [bug#66573] [PATCH v2 09/11] gnu: psm2: Update to 12.0., Ludovic Courtès, 2023/10/19
- [bug#66573] [PATCH v2.1 01/11] gnu: opensm: Use the right version string for the doc directory., Ludovic Courtès, 2023/10/24