memfd: add memory mapping support

* During checkpoint, we add a vma flags: VMA_AREA_MEMFD to denote memfd
  regions.
* Even though memfd is backed by the shmem device, we use the file
  semantics of memfd (via /proc/map_files/<vma>) which we already have
  support for.

Signed-off-by: Nicolas Viennot <Nicolas.Viennot@twosigma.com>

criu/cr-dump.c

@@ -80,6 +80,7 @@
 #include "fault-injection.h"
 #include "dump.h"
 #include "eventpoll.h"
+#include "memfd.h"
 
 /*
  * Architectures can overwrite this function to restore register sets that
@@ -414,7 +415,10 @@ static int dump_filemap(struct vma_area *vma_area, int fd)
 
 	/* Flags will be set during restore in open_filmap() */
 
-	ret = dump_one_reg_file_cond(fd, &id, &p);
+	if (vma->status & VMA_AREA_MEMFD)
+		ret = dump_one_memfd_cond(fd, &id, &p);
+	else
+		ret = dump_one_reg_file_cond(fd, &id, &p);
 
 	vma->shmid = id;
 	return ret;

criu/files-reg.c

@@ -35,6 +35,7 @@
 #include "pstree.h"
 #include "fault-injection.h"
 #include "external.h"
+#include "memfd.h"
 
 #include "protobuf.h"
 #include "util.h"
@@ -1879,7 +1880,10 @@ static int open_filemap(int pid, struct vma_area *vma)
 	flags = vma->e->fdflags;
 
 	if (ctx.flags != flags || ctx.desc != vma->vmfd) {
-		ret = open_path(vma->vmfd, do_open_reg_noseek_flags, &flags);
+		if (vma->e->status & VMA_AREA_MEMFD)
+			ret = memfd_open(vma->vmfd, &flags);
+		else
+			ret = open_path(vma->vmfd, do_open_reg_noseek_flags, &flags);
 		if (ret < 0)
 			return ret;
 
@@ -1909,7 +1913,10 @@ int collect_filemap(struct vma_area *vma)
 			vma->e->fdflags = O_RDONLY;
 	}
 
-	fd = collect_special_file(vma->e->shmid);
+	if (vma->e->status & VMA_AREA_MEMFD)
+		fd = collect_memfd(vma->e->shmid);
+	else
+		fd = collect_special_file(vma->e->shmid);
 	if (!fd)
 		return -1;
 

criu/include/image.h

@@ -83,6 +83,7 @@
 #define VMA_AREA_SOCKET		(1 <<  11)
 #define VMA_AREA_VVAR		(1 <<  12)
 #define VMA_AREA_AIORING	(1 <<  13)
+#define VMA_AREA_MEMFD		(1 <<  14)
 
 #define VMA_CLOSE		(1 <<  28)
 #define VMA_NO_PROT_WRITE	(1 <<  29)

criu/include/memfd.h

@@ -5,10 +5,16 @@
 #include "int.h"
 #include "common/config.h"
 
+struct fd_parms;
+struct file_desc;
+
 extern int is_memfd(dev_t dev, const char *path);
+extern int dump_one_memfd_cond(int lfd, u32 *id, struct fd_parms *parms);
 extern const struct fdtype_ops memfd_dump_ops;
 
+extern int memfd_open(struct file_desc *d, u32 *fdflags);
 extern struct collect_image_info memfd_cinfo;
+extern struct file_desc *collect_memfd(u32 id);
 
 #ifdef CONFIG_HAS_MEMFD_CREATE
 # include <sys/mman.h>

criu/memfd.c

@@ -164,6 +164,13 @@ static int dump_one_memfd(int lfd, u32 id, const struct fd_parms *p)
 	return pb_write_one(img_from_set(glob_imgset, CR_FD_FILES), &fe, PB_FILE);
 }
 
+int dump_one_memfd_cond(int lfd, u32 *id, struct fd_parms *parms)
+{
+	if (fd_id_generate_special(parms, id))
+		return dump_one_memfd(lfd, *id, parms);
+	return 0;
+}
+
 const struct fdtype_ops memfd_dump_ops = {
 	.type		= FD_TYPES__MEMFD,
 	.dump		= dump_one_memfd,
@@ -265,7 +272,7 @@ static int memfd_open_inode(struct memfd_inode *inode)
 	return fd;
 }
 
-static int memfd_open(struct file_desc *d, u32 *fdflags)
+int memfd_open(struct file_desc *d, u32 *fdflags)
 {
 	char lpath[PSFDS];
 	struct memfd_info *mfi;
@@ -384,3 +391,13 @@ struct collect_image_info memfd_cinfo = {
 	.priv_size = sizeof(struct memfd_info),
 	.collect = collect_one_memfd,
 };
+
+struct file_desc *collect_memfd(u32 id) {
+	struct file_desc *fdesc;
+
+	fdesc = find_file_desc_raw(FD_TYPES__MEMFD, id);
+	if (fdesc == NULL)
+		pr_err("No entry for memfd %#x\n", id);
+
+	return fdesc;
+}

criu/proc_parse.c

@@ -41,6 +41,7 @@
 #include "timerfd.h"
 #include "path.h"
 #include "fault-injection.h"
+#include "memfd.h"
 
 #include "protobuf.h"
 #include "images/fdinfo.pb-c.h"
@@ -303,6 +304,26 @@ static int vma_get_mapfile_user(const char *fname, struct vma_area *vma,
 	}
 
 	vfi_dev = makedev(vfi->dev_maj, vfi->dev_min);
+
+	if (is_memfd(vfi_dev, fname)) {
+		struct fd_link link;
+		link.len = strlen(fname);
+		strlcpy(link.name, fname, sizeof(link.name));
+		strip_deleted(&link);
+
+		/*
+		 * The error EPERM will be shown in the following pr_perror().
+		 * It comes from the previous open() call.
+		 */
+		pr_perror("Can't open mapped [%s]", link.name);
+
+		/*
+		 * TODO Perhaps we could do better than failing and dump the
+		 * memory like what is being done in shmem.c
+		 */
+		return -1;
+	}
+
 	if (is_anon_shmem_map(vfi_dev)) {
 		if (!(vma->e->flags & MAP_SHARED))
 			return -1;
@@ -578,7 +599,20 @@ static int handle_vma(pid_t pid, struct vma_area *vma_area,
 		/*
 		 * /dev/zero stands for anon-shared mapping
 		 * otherwise it's some file mapping.
+		 *
+		 * We treat memfd mappings as regular file mappings because
+		 * their backing can be seen as files, which is easy to
+		 * support. So even though memfd is an anonymous shmem, we
+		 * treat it differently.
+		 * Note: maybe we should revisit this as /proc/map_files/<vma>
+		 * may not always be accessible.
 		 */
+
+		if (is_memfd(st_buf->st_dev, file_path)) {
+			vma_area->e->status |= VMA_AREA_MEMFD;
+			goto normal_file;
+		}
+
 		if (is_anon_shmem_map(st_buf->st_dev)) {
 			if (!(vma_area->e->flags & MAP_SHARED))
 				goto err_bogus_mapping;
@@ -594,6 +628,7 @@ static int handle_vma(pid_t pid, struct vma_area *vma_area,
 					vma_area->e->shmid += FI_HUGE_ANON_SHMEM_ID_BASE;
 			}
 		} else {
+normal_file:
 			if (vma_area->e->flags & MAP_PRIVATE)
 				vma_area->e->status |= VMA_FILE_PRIVATE;
 			else