plugin/amdgpu: Support for open dmabuf handles

Add handling for open dmabuf fds. These are file descriptors
meant to be imported by a driver to create shared memory.
Usually, amdgpu runtimes close these right after importing them,
but it's possible that a CRIU process might catch them in the
moment when they exist.

Signed-off-by: David Francis <David.Francis@amd.com>

Author: cfreeamd

criu/files-ext.c

@@ -87,14 +87,14 @@ struct collect_image_info ext_file_cinfo = {
 	.collect = collect_one_ext,
 };
 
-int dump_unsupp_fd(struct fd_parms *p, int lfd, char *more, char *info, FdinfoEntry *e)
+int dump_unsupp_fd(struct fd_parms *p, int lfd, char *more, char *info, FdinfoEntry *e, bool force)
 {
 	int ret;
 
-	ret = do_dump_gen_file(p, lfd, &ext_dump_ops, e);
+	ret = do_dump_gen_file(p, lfd, &ext_dump_ops, e, force);
 	if (ret == 0)
 		return 0;
 	if (ret == -ENOTSUP)
 		pr_err("Can't dump file %d of that type [%o] (%s %s)\n", p->fd, p->stat.st_mode, more, info);
-	return -1;
+	return ret;
 }

criu/files.c

@@ -329,7 +329,8 @@ uint32_t make_gen_id(uint32_t st_dev, uint32_t st_ino, uint64_t pos)
 	return st_dev ^ st_ino ^ pos_hi ^ pos_low;
 }
 
-int do_dump_gen_file(struct fd_parms *p, int lfd, const struct fdtype_ops *ops, FdinfoEntry *e)
+int do_dump_gen_file(struct fd_parms *p, int lfd,
+					const struct fdtype_ops *ops, FdinfoEntry *e, bool force)
 {
 	int ret = -1;
 
@@ -339,7 +340,7 @@ int do_dump_gen_file(struct fd_parms *p, int lfd, const struct fdtype_ops *ops,
 	e->flags = p->fd_flags;
 
 	ret = fd_id_generate(p->pid, e, p);
-	if (ret == 1) /* new ID generated */
+	if (ret == 1 || force) /* new ID generated */
 		ret = ops->dump(lfd, e->id, p);
 	else
 		/* Remove locks generated by the fd before going to the next */
@@ -484,19 +485,19 @@ static int dump_chrdev(struct fd_parms *p, int lfd, FdinfoEntry *e)
 		}
 
 		sprintf(more, "%d:%d", maj, minor(p->stat.st_rdev));
-		err = dump_unsupp_fd(p, lfd, "chr", more, e);
+		err = dump_unsupp_fd(p, lfd, "chr", more, e, false);
 		p->link = link_old;
 		return err;
 	}
 	}
 
-	err = do_dump_gen_file(p, lfd, ops, e);
+	err = do_dump_gen_file(p, lfd, ops, e, false);
 	p->link = link_old;
 	return err;
 }
 
 static int dump_one_file(struct pid *pid, int fd, int lfd, struct fd_opts *opts, struct parasite_ctl *ctl,
-			 FdinfoEntry *e, struct parasite_drain_fd *dfds)
+			 FdinfoEntry *e, struct parasite_drain_fd *dfds, bool force)
 {
 	struct fd_parms p = FD_PARMS_INIT;
 	const struct fdtype_ops *ops;
@@ -552,14 +553,14 @@ static int dump_one_file(struct pid *pid, int fd, int lfd, struct fd_opts *opts,
 			ops = &bpfmap_dump_ops;
 #endif
 		else
-			return dump_unsupp_fd(&p, lfd, "anon", link, e);
+			return dump_unsupp_fd(&p, lfd, "anon", link, e, force);
 
-		return do_dump_gen_file(&p, lfd, ops, e);
+		return do_dump_gen_file(&p, lfd, ops, e, force);
 	}
 
 	if (p.fs_type == PID_FS_MAGIC) {
 		ops = &pidfd_dump_ops;
-		return do_dump_gen_file(&p, lfd, ops, e);
+		return do_dump_gen_file(&p, lfd, ops, e, force);
 	}
 
 	if (S_ISREG(p.stat.st_mode) || S_ISDIR(p.stat.st_mode) || S_ISLNK(p.stat.st_mode)) {
@@ -576,9 +577,9 @@ static int dump_one_file(struct pid *pid, int fd, int lfd, struct fd_opts *opts,
 		else if (check_ns_proc(&link))
 			ops = &nsfile_dump_ops;
 		else
-			return dump_unsupp_fd(&p, lfd, "reg", link.name + 1, e);
+			return dump_unsupp_fd(&p, lfd, "reg", link.name + 1, e, force);
 
-		return do_dump_gen_file(&p, lfd, ops, e);
+		return do_dump_gen_file(&p, lfd, ops, e, force);
 	}
 
 	if (S_ISFIFO(p.stat.st_mode)) {
@@ -587,7 +588,7 @@ static int dump_one_file(struct pid *pid, int fd, int lfd, struct fd_opts *opts,
 		else
 			ops = &fifo_dump_ops;
 
-		return do_dump_gen_file(&p, lfd, ops, e);
+		return do_dump_gen_file(&p, lfd, ops, e, force);
 	}
 
 	/*
@@ -598,7 +599,7 @@ static int dump_one_file(struct pid *pid, int fd, int lfd, struct fd_opts *opts,
 	if (fill_fdlink(lfd, &p, &link))
 		memzero(&link, sizeof(link));
 
-	return dump_unsupp_fd(&p, lfd, "unknown", link.name + 1, e);
+	return dump_unsupp_fd(&p, lfd, "unknown", link.name + 1, e, force);
 }
 
 int dump_my_file(int lfd, u32 *id, int *type)
@@ -610,7 +611,7 @@ int dump_my_file(int lfd, u32 *id, int *type)
 	me.real = getpid();
 	me.ns[0].virt = -1; /* FIXME */
 
-	if (dump_one_file(&me, lfd, lfd, &fdo, NULL, &e, NULL))
+	if (dump_one_file(&me, lfd, lfd, &fdo, NULL, &e, NULL, false))
 		return -1;
 
 	*id = e.id;
@@ -625,6 +626,8 @@ int dump_task_files_seized(struct parasite_ctl *ctl, struct pstree_item *item, s
 	struct fd_opts *opts = NULL;
 	int i, ret = -1;
 	int off, nr_fds = min((int)PARASITE_MAX_FDS, dfds->nr_fds);
+	int *retry_indices = NULL;
+	int retry_count = 0;
 
 	pr_info("\n");
 	pr_info("Dumping opened files (pid: %d)\n", item->pid->real);
@@ -642,6 +645,10 @@ int dump_task_files_seized(struct parasite_ctl *ctl, struct pstree_item *item, s
 	if (!img)
 		goto err;
 
+	retry_indices = xmalloc(dfds->nr_fds * sizeof(int)); /* Allocate memory for retry indices*/
+	if (!retry_indices)
+		goto err;
+
 	ret = 0; /* Don't fail if nr_fds == 0 */
 	for (off = 0; ret == 0 && off < dfds->nr_fds; off += nr_fds) {
 		if (nr_fds + off > dfds->nr_fds)
@@ -654,9 +661,32 @@ int dump_task_files_seized(struct parasite_ctl *ctl, struct pstree_item *item, s
 		for (i = 0; i < nr_fds; i++) {
 			FdinfoEntry e = FDINFO_ENTRY__INIT;
 
-			ret = dump_one_file(item->pid, dfds->fds[i + off], lfds[i], opts + i, ctl, &e, dfds);
+			ret = dump_one_file(item->pid, dfds->fds[i + off], lfds[i], opts + i, ctl, &e, dfds, false);
+			if (ret == -EAGAIN) {
+				retry_indices[retry_count++] = i;
+				ret = 0; //* Reset ret to continue the loop */
+				continue;
+			} else if (ret)
+				break;
+
+			ret = pb_write_one(img, &e, PB_FDINFO);
 			if (ret)
 				break;
+		}
+		/* Dmabuf FDs cannot be dumped unless we have first dumped the device file that exported them.
+		 * For that reason, retry dump of files that failed to dump the first time. Specifying force
+		 * here makes the dump go through even if a file id has already been assigned.
+		 */
+		for (i = 0; i < retry_count; i++) {
+			int idx = retry_indices[i];
+			FdinfoEntry e = FDINFO_ENTRY__INIT;
+
+			ret = dump_one_file(item->pid, dfds->fds[idx + off], lfds[idx],
+											opts + idx, ctl, &e, dfds, true);
+			if (ret) {
+				pr_err("Retry failed for fd index %d\n", idx);
+				break;
+			}
 
 			ret = pb_write_one(img, &e, PB_FDINFO);
 			if (ret)
@@ -673,6 +703,8 @@ int dump_task_files_seized(struct parasite_ctl *ctl, struct pstree_item *item, s
 		close_image(img);
 	xfree(opts);
 	xfree(lfds);
+	xfree(retry_indices);
+
 	return ret;
 }
 

criu/include/files.h

@@ -138,7 +138,8 @@ struct fdtype_ops {
 struct cr_img;
 
 extern int dump_my_file(int lfd, u32 *, int *type);
-extern int do_dump_gen_file(struct fd_parms *p, int lfd, const struct fdtype_ops *ops, FdinfoEntry *e);
+extern int do_dump_gen_file(struct fd_parms *p, int lfd,
+					const struct fdtype_ops *ops, FdinfoEntry *e, bool force);
 struct parasite_drain_fd;
 int dump_task_files_seized(struct parasite_ctl *ctl, struct pstree_item *item, struct parasite_drain_fd *dfds);
 int predump_task_files(int pid);
@@ -177,7 +178,8 @@ extern int close_old_fds(void);
 extern int shared_fdt_prepare(struct pstree_item *item);
 
 extern struct collect_image_info ext_file_cinfo;
-extern int dump_unsupp_fd(struct fd_parms *p, int lfd, char *more, char *info, FdinfoEntry *);
+extern int dump_unsupp_fd(struct fd_parms *p, int lfd, char *more,
+								char *info, FdinfoEntry *, bool force);
 
 extern int inherit_fd_parse(char *optarg);
 extern int inherit_fd_add(int fd, char *key);

criu/sockets.c

@@ -766,7 +766,7 @@ int dump_socket(struct fd_parms *p, int lfd, FdinfoEntry *e)
 		return -1;
 	}
 
-	return do_dump_gen_file(p, lfd, ops, e);
+	return do_dump_gen_file(p, lfd, ops, e, false);
 }
 
 static int inet_receive_one(struct nlmsghdr *h, struct ns_id *ns, void *arg)

plugins/amdgpu/Makefile

@@ -27,7 +27,7 @@ endif
 criu-amdgpu.pb-c.c: criu-amdgpu.proto
 		protoc --proto_path=. --c_out=. criu-amdgpu.proto
 
-amdgpu_plugin.so: amdgpu_plugin.c amdgpu_plugin_drm.c amdgpu_plugin_topology.c amdgpu_plugin_util.c criu-amdgpu.pb-c.c amdgpu_socket_utils.c
+amdgpu_plugin.so: amdgpu_plugin.c amdgpu_plugin_drm.c amdgpu_plugin_dmabuf.c amdgpu_plugin_topology.c amdgpu_plugin_util.c criu-amdgpu.pb-c.c amdgpu_socket_utils.c
 	$(CC) $(PLUGIN_CFLAGS) $(shell $(COMPEL) includes) $^ -o $@ $(PLUGIN_INCLUDE) $(PLUGIN_LDFLAGS) $(LIBDRM_INC)
 
 amdgpu_plugin_clean:

plugins/amdgpu/amdgpu_plugin.c

@@ -38,6 +38,7 @@
 #include "rst-malloc.h"
 
 #include "common/list.h"
+#include "amdgpu_plugin_dmabuf.h"
 #include "amdgpu_plugin_drm.h"
 #include "amdgpu_plugin_util.h"
 #include "amdgpu_plugin_topology.h"
@@ -46,7 +47,7 @@
 #include "img-streamer.h"
 #include "image.h"
 #include "cr_options.h"
-
+#include "util.h"
 struct vma_metadata {
 	struct list_head list;
 	uint64_t old_pgoff;
@@ -1403,7 +1404,17 @@ int amdgpu_plugin_dump_file(int fd, int id)
 		return -1;
 	}
 
-	/* Check whether this plugin was called for kfd or render nodes */
+	/* Check whether this plugin was called for kfd, dmabuf or render nodes */
+	ret = get_dmabuf_info(fd, &st);
+	if (ret < 0) {
+		pr_perror("Failed to get dmabuf info");
+		return -1;
+	} else if (ret == 0) {
+		pr_info("Dumping dmabuf fd = %d\n", fd);
+		ret = amdgpu_plugin_dmabuf_dump(fd, id, &st);
+		return ret;
+	}
+
 	if (major(st.st_rdev) != major(st_kfd.st_rdev) || minor(st.st_rdev) != 0) {
 
 		/* This is RenderD dumper plugin, for now just save renderD
@@ -1415,9 +1426,6 @@ int amdgpu_plugin_dump_file(int fd, int id)
 			return ret;
 
 		ret = record_dumped_fd(fd, true);
-		if (ret)
-			return ret;
-
 		/* Need to return success here so that criu can call plugins for renderD nodes */
 		return ret;
 	}
@@ -1541,7 +1549,6 @@ static int restore_devices(struct kfd_ioctl_criu_args *args, CriuKfd *e)
 	int ret = 0, bucket_index = 0;
 
 	pr_debug("Restoring %d devices\n", e->num_of_gpus);
-
 	args->num_devices = e->num_of_gpus;
 	device_buckets = xzalloc(sizeof(*device_buckets) * args->num_devices);
 	if (!device_buckets)
@@ -1826,12 +1833,17 @@ int amdgpu_plugin_restore_file(int id, bool *retry_needed)
 		 * first as we assume restore_maps is already filled. Need to fix this later.
 		 */
 		snprintf(img_path, sizeof(img_path), IMG_DRM_FILE, id);
-		pr_info("Restoring RenderD %s\n", img_path);
 
 		img_fp = open_img_file(img_path, false, &img_size);
-		if (!img_fp)
-			return -EINVAL;
-
+		if (!img_fp) {
+			ret = amdgpu_plugin_dmabuf_restore(id);
+			if (ret == 1) {
+				*retry_needed = true;
+				return 0;
+			}
+			return ret;
+		}
+		pr_info("Restoring RenderD %s\n", img_path);
 		pr_debug("RenderD Image file size:%ld\n", img_size);
 		buf = xmalloc(img_size);
 		if (!buf) {