Support 32-bit architectures #11028

Gonzalosilvalde · 2025-05-15T16:03:13Z

Summary

This Merge Request enhances libfabric’s portability on 32-bit systems by:

Replacing hard-coded format specifiers (%lx, %zx, etc.) with the corresponding <inttypes.h> macros (PRIuPTR, PRIxPTR, PRIx64, PRIu64, etc.).
Adjusting pointer and integer casts to use uintptr_t instead of fixed-width types (e.g. uint64_t) for address arithmetic.
Introducing size_t temporaries when consuming I/O vectors to ensure correct sizing on all architectures.
Format specifier correction: Fixed error messages that used %lu for size_t, replacing them with %zu.

Types of changes

- "buf %p len %zu addr %zu data %lu " \
- "flags 0x%zx ctx %p\n", \
- buf, len, addr, (uint64_t)data, \
- (uint64_t)flags, context); \
+ "buf %p len %zu addr %" PRIuPTR " data 0x%" PRIx64 " " \
+ "flags 0x%" PRIxPTR " ctx %p\n", \
+ (void *)(buf), (size_t)(len), (uintptr_t)(addr), (uint64_t)(data), \
+ (uintptr_t)flags, (void *)(context)); \

- ofi_hmem_dev_unregister(attr.iface, (uint64_t)attr.hmem_data);
+ ofi_hmem_dev_unregister(attr.iface, (uintptr_t)attr.hmem_data);

References

Github issue libfabric 1.20.1 fails to build on Fedora Rahide i686 #9763 (libfabric 1.20.1 fails to build on Fedora Rahide i686 #9763)
Debian bug #1102068 (https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1102068)
This Merge request is based on the Salsa patch (https://salsa.debian.org/hpc-team/libfabric/-/commit/4260159f6cccb80113c85eb909d8f759122710fe)

j-xiong · 2025-05-15T16:57:07Z

include/ofi_cma.h

+		size_t local_size = local_cnt;
+		size_t remote_size = remote_cnt;
+
+		ofi_consume_iov(local, &local_size, (size_t)ret);
+		ofi_consume_iov(remote, &remote_size, (size_t)ret);
+
+		local_cnt = local_size;
+		remote_cnt = remote_size;


should change the type of local_cnt and remote_cnt instead.

@aingerson The type of the parameters passed in for this call is actually size_t.

j-xiong · 2025-05-15T17:01:52Z

prov/hook/hook_hmem/src/hook_hmem.c

 		      FI_REMOTE_READ | FI_REMOTE_WRITE;
 	attr.offset = 0;
-	attr.requested_key = (uint64_t) *hmem_desc;
+	attr.requested_key = (uintptr_t) *hmem_desc;


Here we are indeed expecting uint64_t. The conversion works for both 32-bit and 64-bit machines is (uint64_t)(uintptr_t) *hmem_desc

j-xiong · 2025-05-15T17:08:57Z

prov/lnx/include/lnx.h

 	rc = ofi_mr_cache_search(&lnx_dom->ld_mr_cache, &info, mre);
 	if (rc) {
-		ofi_hmem_dev_unregister(attr.iface, (uint64_t)attr.hmem_data);
+		ofi_hmem_dev_unregister(attr.iface, (uintptr_t)attr.hmem_data);


same here, use (uint64_t)(uintptr_t).

j-xiong · 2025-05-15T17:09:23Z

prov/lnx/src/lnx_domain.c

 	struct ofi_mr *mr = (struct ofi_mr *)entry->data;

-	ofi_hmem_dev_unregister(mr->iface, (uint64_t) mr->hmem_data);
+	ofi_hmem_dev_unregister(mr->iface, (uintptr_t) mr->hmem_data);


j-xiong · 2025-05-15T17:12:02Z

prov/mrail/src/mrail_domain.c

 	memcpy(mr_map, map->raw_key, map->key_size);

-	*(map->key) = (uint64_t)mr_map;
+	*(map->key) = (uintptr_t)mr_map;


j-xiong · 2025-05-15T17:37:53Z

prov/verbs/src/verbs_init.c

-		 vrb_gl_data.dmabuf_support ? "enabled" : "disabled");
+			vrb_gl_data.dmabuf_support ? "enabled" : "disabled");


j-xiong · 2025-05-15T17:38:10Z

prov/verbs/src/verbs_init.c

-	if (vrb_get_param_str("iface", "The prefix or the full name of the "
-			      "network interface associated with the verbs "
-			      "device", &vrb_gl_data.iface)) {
+	if (vrb_get_param_str("iface",
+			"Network interface prefix or full name",
+			&vrb_gl_data.iface)) {


j-xiong · 2025-05-15T17:38:23Z

prov/verbs/src/verbs_init.c

-	if (vrb_get_param_bool("dgram_use_name_server", "The option that "
-			       "enables/disables OFI Name Server thread used "
-			       "to resolve IP-addresses to provider specific "
-			       "addresses. If MPI is used, the NS is disabled "
-			       "by default.", &vrb_gl_data.dgram.use_name_server)) {
+
+	if (vrb_read_param_bool("dgram_use_name_server",
+			"Enable/disable OFI Name Server thread",
+			&vrb_gl_data.dgram.use_name_server)) {


j-xiong · 2025-05-15T17:38:35Z

prov/verbs/src/verbs_init.c

-	if (vrb_get_param_int("dgram_name_server_port", "The port on which "
-			      "the name server thread listens incoming "
-			      "requests.", &vrb_gl_data.dgram.name_server_port) ||
-	    (vrb_gl_data.dgram.name_server_port < 0 ||
-	     vrb_gl_data.dgram.name_server_port > 65535)) {
+
+	if (vrb_get_param_int("dgram_name_server_port",
+			"Port for name server thread",
+			&vrb_gl_data.dgram.name_server_port) ||
+			vrb_gl_data.dgram.name_server_port < 0 ||
+			vrb_gl_data.dgram.name_server_port > 65535) {


j-xiong · 2025-05-15T17:39:49Z

prov/verbs/src/verbs_mr.c

-	mr = ibv_reg_dmabuf_mr(pd, offset, len, (uint64_t)buf/* iova */,
-			       fd, vrb_access);
+	{
+		uintptr_t buf_addr = (uintptr_t)buf;
+		uint64_t iova = (uint64_t)buf_addr;
+		mr = ibv_reg_dmabuf_mr(pd, offset, len, iova,
+								fd, vrb_access);
+	}
+


Can simply cast buf to (uint64_t)(uintptr_t)

j-xiong · 2025-05-15T17:42:15Z

Also please sign off your commit.

j-xiong

Looks mostly good except for a few formatting issues.

For the DCO error: Commit sha: c9b9374, Author: Gonzalosilvalde, Committer: Gonzalosilvalde; Expected "Gonzalosilvalde gonzalo.silvalde@udc.es", but got "Gonzalo Silvalde gonzalo.silvalde@gmail.com". It seems that the machine you used to create the commit has git email configured differently from the one used in "Signed-off-by".

j-xiong · 2025-05-27T15:24:10Z

prov/rxm/src/rxm_cq.c

 		void *res_buf = resp_hdr->data + offset;
-		void *dst_buf = (void *) req_hdr->rma_ioc[i].addr;
-
+		void *dst_buf = (void *)(uintptr_t) req_hdr->rma_ioc[i].addr;


Please keep the empty line here.

j-xiong · 2025-05-27T15:26:26Z

prov/verbs/src/verbs_mr.c

-	mr = ibv_reg_dmabuf_mr(pd, offset, len, (uint64_t)buf/* iova */,
-			       fd, vrb_access);
+	mr = ibv_reg_dmabuf_mr(pd, offset, len, (uint64_t)(uintptr_t)buf/* iova */,
+						   fd, vrb_access);


The alignment of L88 is off.

j-xiong · 2025-05-29T03:02:48Z

Any changed files under prov/sm2 need to be run through clang-format to pass the format check. That is for the sm provider only.

Gonzalosilvalde · 2025-06-10T10:41:51Z

Sorry for the delay in making these small updates, I was on vacation without a computer handy. I see there are some merge conflicts, I’m looking into exactly what’s causing them now and will resolve the issues shortly. Apologies for the inconvenience, and thanks for your patience.

Signed-off-by: Gonzalo Silvalde <gonzalo.silvalde@gmail.com>

Gonzalosilvalde · 2026-01-26T10:48:39Z

Good afternoon, @j-xiong . After a couple of months, I've come back to this problem, but it's still failing. I thought that the failures related to Jenkins and AWS CI had to do with fabtest not being adapted to 32 bits, so I made some changes, but I see that this isn't the problem and the errors persist. I don't know if you could give me some guidance on what might be failing. Best regards.

j-xiong requested changes May 15, 2025

View reviewed changes

Gonzalosilvalde force-pushed the 32-bit branch from eba26b7 to c9b9374 Compare May 24, 2025 21:01

j-xiong reviewed May 27, 2025

View reviewed changes

Gonzalosilvalde force-pushed the 32-bit branch from c9b9374 to 429453f Compare June 10, 2025 10:36

Gonzalosilvalde force-pushed the 32-bit branch 3 times, most recently from 5ecd575 to 017d68e Compare June 10, 2025 16:33

lowpolyneko mentioned this pull request Sep 26, 2025

Current embedded libfabric version fails to build on i386 on newer GCC versions pmodels/mpich#7599

Open

Gonzalosilvalde force-pushed the 32-bit branch 4 times, most recently from ae34d1b to 22a25d6 Compare January 25, 2026 22:36

enable 32-bit build support

d3997a1

Signed-off-by: Gonzalo Silvalde <gonzalo.silvalde@gmail.com>

Gonzalosilvalde force-pushed the 32-bit branch from 22a25d6 to d3997a1 Compare January 26, 2026 06:26

		vrb_gl_data.dmabuf_support ? "enabled" : "disabled");
		vrb_gl_data.dmabuf_support ? "enabled" : "disabled");

Support 32-bit architectures #11028

Are you sure you want to change the base?

Support 32-bit architectures #11028

Conversation

Gonzalosilvalde commented May 15, 2025

Summary

Types of changes

References

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

j-xiong commented May 15, 2025

Uh oh!

j-xiong left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

j-xiong commented May 29, 2025

Uh oh!

Gonzalosilvalde commented Jun 10, 2025

Uh oh!

Gonzalosilvalde commented Jan 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants