criu/0027-mmap-restore-dev-hisi_sec2-deivce-vma.patch

From fe19a2639373175c134fa51a7c1c26ca5306d22c Mon Sep 17 00:00:00 2001
From: "fu.lin" <fulin10@huawei.com>
Date: Fri, 10 Sep 2021 16:06:55 +0800
Subject: [PATCH 27/72] mmap: restore /dev/hisi_sec2* deivce vma

There are two kinds of vmas: anonymous vma and file-based vma. For
anonymous vma, criu just map area and fill content to it; for file-based
vma, criu preprocess it, such as setting `open_vm()` callback function.

`/dev/hisi_sec2*` char device is different from the normal. The `open`,
`mmap`, and `close` syscall actions has a special meaning.
  - `open`: allocate physical resource of the device
  - `mmap`: create instance
  - `close`: release physical resource
The vma means the instance in this device. One fd may be associated with
a group instances: one mmio (vma size is 2 pages, pgoff is 0), one dus
(vma size is 37 pages, pgoff is 0x2000). As for dus vma, it's split two
vmas by `mprotect(addr, 0x5000, PROT_READ)`: one size is 0x20000, one
size is 0x5000.

This patch makes the /dev/hisi_sec* restore possible. Idea:
  It's impossible for criu to know the relationship between vma and the
mapped file fd. Therefore, just collect the total fds number during
collecting /dev/hisi_sec* files, then the fd is tagged that which
function is used during vma restoration, and aissign the unused fd to the
specific vma. And during `mmap()` process, dus vma is splitted by `mprotect`.

Note:
- criu use ino to index the fd.
- this physical device drivers is hisi_sec2.ko, which is located in
  `drivers/crypto/hisilicon/sec2/` of linux kernel.
- this device name has prefix "hisi_sec2" that is found from
  `drivers/crypto/hisilicon/sec2/sec_main.c`.

Conflict:NA
Reference:https://gitee.com/src-openeuler/criu/pulls/21
Signed-off-by: fu.lin <fulin10@huawei.com>
---
 criu/files-chr.c         | 130 +++++++++++++++++++++++++++++++++++++--
 criu/include/files-chr.h |  16 +++++
 criu/include/vma.h       |  12 ++++
 criu/pie/restorer.c      | 130 ++++++++++++++++++++++++++++++++++++++-
 criu/proc_parse.c        |   4 +-
 5 files changed, 284 insertions(+), 8 deletions(-)

diff --git a/criu/files-chr.c b/criu/files-chr.c
index 315e9c6..95d93e1 100644
--- a/criu/files-chr.c
+++ b/criu/files-chr.c
@@ -6,6 +6,9 @@
 #include "log.h"
 
 #include "protobuf.h"
+#include "rst-malloc.h"
+
+static unsigned hisi_sec_fds_n;
 
 /* Checks if file descriptor @lfd is infinibandevent */
 int is_infiniband_link(char *link)
@@ -16,11 +19,14 @@ int is_infiniband_link(char *link)
 static int chrfile_open(struct file_desc *d, int *new_fd)
 {
 	int fd, mntns_root;
-	int ret = 0;
+	int ret = -1;
 	struct chrfile_info *ci;
 
 	ci = container_of(d, struct chrfile_info, d);
 
+	pr_info("charfile: Opening %s (repair %d index %d)\n",
+			ci->path, ci->cfe->repair, ci->cfe->index);
+
 	if (ci->cfe->repair)
 		ci->cfe->flags |= O_REPAIR;
 
@@ -32,6 +38,7 @@ static int chrfile_open(struct file_desc *d, int *new_fd)
 	}
 
 	*new_fd = fd;
+	ret = 0;
 
 	return ret;
 }
@@ -52,10 +59,12 @@ static int collect_one_chrfile(void *o, ProtobufCMessage *base, struct cr_img *i
 	else
 		ci->path = ci->cfe->name;
 
-	pr_info("Collected chr file: %#x, name: %s\n", ci->cfe->id, ci->path);
-	file_desc_add(&ci->d, ci->cfe->id, &chrfile_desc_ops);
+	/* collect `/dev/hisi_sec2*` fds */
+	if (strstr(ci->path, HISI_SEC_DEV) != NULL)
+		hisi_sec_fds_n += 1;
 
-	return 0;
+	pr_info("Collected chr file: %#x, name: %s\n", ci->cfe->id, ci->path);
+	return file_desc_add(&ci->d, ci->cfe->id, &chrfile_desc_ops);
 }
 
 struct collect_image_info chrfile_cinfo = {
@@ -65,6 +74,7 @@ struct collect_image_info chrfile_cinfo = {
 	.collect = collect_one_chrfile,
 };
 
+static int handle_hisi_vma(struct list_head *fds, struct vma_area *vma);
 int collect_chr_map(struct pstree_item *me, struct vma_area *vma)
 {
 	struct list_head *list = &rsti(me)->fds;
@@ -72,6 +82,12 @@ int collect_chr_map(struct pstree_item *me, struct vma_area *vma)
 	struct chrfile_info *ci;
 	bool exist_fd;
 
+	if (strstr(vma->e->name, HISI_SEC_DEV) != NULL) {
+		if (handle_hisi_vma(list, vma) != 0) {
+			return -1;
+		} else
+			goto out;
+	}
 
 	list_for_each_entry_safe(fle, tmp, list, ps_list) {
 		struct file_desc *d = fle->desc;
@@ -91,5 +107,111 @@ int collect_chr_map(struct pstree_item *me, struct vma_area *vma)
 	if (!exist_fd)
 		return -EEXIST;
 
+out:
+	pr_info("  `- find fd %ld for dev %s at this vma\n", vma->e->fd, vma->e->name);
+
+	return 0;
+}
+
+#define MAX_HISI_SEC_SIZE	3 /* one physical device expose three char dev */
+static struct hlist_head hisi_sec_fds_hash[MAX_HISI_SEC_SIZE];
+
+static int collect_hisi_sec_fds(struct list_head *list)
+{
+	struct fdinfo_list_entry *fle, *tmp;
+	struct chrfile_info *ci;
+	struct file_desc *d;
+	struct hisi_sec_desc *desc;
+	int idx;
+	int nr = 0;
+
+	for (idx = 0; idx < MAX_HISI_SEC_SIZE; idx++)
+		INIT_HLIST_HEAD(&hisi_sec_fds_hash[idx]);
+
+	list_for_each_entry_safe(fle, tmp, list, ps_list) {
+		d = fle->desc;
+
+		if (d->ops->type != FD_TYPES__CHR)
+			continue;
+
+		ci = container_of(d, struct chrfile_info, d);
+
+		if (strstr(ci->path, HISI_SEC_DEV) != NULL) {
+			desc = shmalloc(sizeof(*desc));
+			if (desc == NULL)
+				return -ENOMEM;
+
+			desc->name = ci->path;
+			desc->fd = fle->fe->fd;
+			desc->mmio = desc->dus = 0;
+
+			idx = (ci->path[strlen(ci->path)-1] - '0') % MAX_HISI_SEC_SIZE;
+			hlist_add_head(&desc->hash, &hisi_sec_fds_hash[idx]);
+
+			nr += 1;
+		}
+	}
+
+	return nr;
+}
+
+static long delivery_hisi_sec_fd(struct list_head *fds, struct vma_area *vma)
+{
+	extern unsigned hisi_sec_fds_n; /* defined in criu/files.c */
+	static bool initialized  = false;
+	struct hisi_sec_desc *desc;
+	int fd = -1, idx;
+
+	if (!initialized) {
+		int nr;
+
+		pr_info("find %d fds for hisi_sec char device\n", hisi_sec_fds_n);
+
+		nr = collect_hisi_sec_fds(fds);
+		if (nr != hisi_sec_fds_n) {
+			pr_err("Collected fds(%d) aren't equal opened(%d)\n",
+			       nr, hisi_sec_fds_n);
+			return -1;
+		}
+
+		initialized = true;
+	} else if (vma->e->pgoff != HISI_SEC_MMIO && vma->e->pgoff != HISI_SEC_DUS) {
+		/* It's impossible value for fd, just as a tag to show it's a
+		 * vma by `mprotect` syscall.
+		 */
+		return LONG_MAX;
+	}
+
+	idx = (vma->e->name[strlen(vma->e->name)-1] - '0') % MAX_HISI_SEC_SIZE;
+	hlist_for_each_entry(desc, &hisi_sec_fds_hash[idx], hash) {
+		if (strcmp(desc->name, vma->e->name) != 0)
+			continue;
+
+		if (vma->e->pgoff == HISI_SEC_MMIO && !desc->mmio) {
+			fd = desc->fd;
+			desc->mmio = true;
+			break;
+		} else if (vma->e->pgoff == HISI_SEC_DUS && !desc->dus) {
+			fd = desc->fd;
+			desc->dus = true;
+			break;
+		}
+	}
+
+	return fd;
+}
+
+static int handle_hisi_vma(struct list_head *fds, struct vma_area *vma)
+{
+	long fd = delivery_hisi_sec_fd(fds, vma);
+
+	if (fd < 0) {
+		pr_err("find fd for char dev vma pgoff %lx named %s failed.\n",
+		       vma->e->pgoff, vma->e->name);
+		return -1;
+	}
+
+	vma->e->fd = fd;
+
 	return 0;
 }
diff --git a/criu/include/files-chr.h b/criu/include/files-chr.h
index 5be11f5..26b8fb2 100644
--- a/criu/include/files-chr.h
+++ b/criu/include/files-chr.h
@@ -22,4 +22,20 @@ bool find_devname(const char *name);
 int collect_chr_map(struct pstree_item *me, struct vma_area *vma);
 int is_infiniband_link(char *link);
 
+struct hisi_sec_desc {
+	struct hlist_node hash;
+	char *name;
+	bool mmio;
+	bool dus;
+	int fd;
+};
+
+#define HISI_SEC_DEV	"hisi_sec2"	/* `/dev/hisi_sec2*` char device */
+
+/* here is the selection of offset in `mmap`, they're from drivers */
+enum hisi_sec_dev {
+	HISI_SEC_MMIO	= 0x0,
+	HISI_SEC_DUS	= 0x2000,
+};
+
 #endif /* __CRIU_FILES_CHR_H__ */
diff --git a/criu/include/vma.h b/criu/include/vma.h
index ed9f31e..2b6e86f 100644
--- a/criu/include/vma.h
+++ b/criu/include/vma.h
@@ -125,4 +125,16 @@ static inline bool vma_entry_can_be_lazy(VmaEntry *e)
 		!(vma_entry_is(e, VMA_AREA_VDSO)) && !(vma_entry_is(e, VMA_AREA_VSYSCALL)));
 }
 
+struct vma_attr {
+	int prot;
+	int flags;
+};
+
+enum ALIEN_MAP_METHOD {
+	PGOFF_IS_ZERO,
+	MAP_THEN_PROTECT,
+
+	MAX_ALIEN_MAP_METHOD,
+};
+
 #endif /* __CR_VMA_H__ */
diff --git a/criu/pie/restorer.c b/criu/pie/restorer.c
index 549bbd6..dcc922e 100644
--- a/criu/pie/restorer.c
+++ b/criu/pie/restorer.c
@@ -37,6 +37,7 @@
 #include "uffd.h"
 #include "sched.h"
 #include "notifier.h"
+#include "files-chr.h"
 
 #include "common/lock.h"
 #include "common/page.h"
@@ -861,6 +862,129 @@ static unsigned long restore_mapping(VmaEntry *vma_entry)
 	return addr;
 }
 
+static unsigned long restore_map_then_protect_mapping(VmaEntry *curr,
+						      struct vma_attr *curr_attr,
+						      VmaEntry *next,
+						      struct vma_attr *next_attr)
+{
+	int retval;
+	unsigned long addr;
+
+	if (next->fd != LONG_MAX
+			|| curr->end != next->start
+			|| (vma_entry_len(curr) + curr->pgoff) != next->pgoff
+			|| curr->prot == next->prot
+			|| curr->flags != next->flags) {
+		pr_err("They looks not currect:\n");
+		pr_err("  `- vma A: (%x %x %d %lx)\n",
+				curr_attr->prot, curr_attr->flags,
+				(int)curr->fd, curr->pgoff);
+		pr_err("  `- vma B: (%x %x %d %lx)\n",
+				next_attr->prot, next_attr->flags,
+				(int)next->fd, next->pgoff);
+		return -1;
+	}
+
+	pr_info("\tmmap(%x %x %d %lx) in map then protect mapping\n",
+				curr_attr->prot, curr_attr->flags,
+				(int)curr->fd, curr->pgoff);
+
+	addr = sys_mmap(decode_pointer(curr->start),
+		 vma_entry_len(curr) + vma_entry_len(next),
+		 curr_attr->prot, curr_attr->flags, curr->fd, curr->pgoff);
+	if (addr != curr->start) {
+		pr_err("%s: mmap failed with code %ld\n", __func__, addr);
+		goto out;
+	}
+
+	pr_info("\t  mprotect(%x)\n", next_attr->prot);
+	retval = sys_mprotect(decode_pointer(next->start),
+			      vma_entry_len(next), next_attr->prot);
+	if (retval != 0) {
+		addr = retval;
+		pr_err("%s: mprotect failed with code %d\n", __func__, retval);
+	}
+
+out:
+	return addr;
+}
+
+static unsigned long restore_pgoff_is_zero_mapping(VmaEntry *curr, struct vma_attr *attr)
+{
+	unsigned long addr;
+
+	pr_debug("\tmmap(%x %x %d %lx) in pgoff is zero mapping\n",
+			attr->prot, attr->flags, (int)curr->fd, curr->pgoff);
+
+	addr = sys_mmap(decode_pointer(curr->start),
+			vma_entry_len(curr),
+			attr->prot, attr->flags,
+			curr->fd, curr->pgoff);
+
+	return addr;
+}
+
+static unsigned long restore_hisi_sec_mapping(struct task_restore_args *args,
+						int i, int *step)
+{
+	VmaEntry *curr = args->vmas + i;
+	VmaEntry *next = args->vmas + i + 1;
+	struct vma_attr curr_attr = {
+		.prot = curr->prot,
+		.flags = curr->flags | MAP_FIXED,
+	};
+	struct vma_attr next_attr = {
+		.prot = next->prot,
+		.flags = next->flags | MAP_FIXED,
+	};
+	unsigned long addr;
+
+	switch (curr->pgoff) {
+	case HISI_SEC_MMIO:
+		addr = restore_pgoff_is_zero_mapping(curr, &curr_attr);
+		break;
+	case HISI_SEC_DUS:
+		*step = 2;
+		addr = restore_map_then_protect_mapping(curr, &curr_attr, next, &next_attr);
+		break;
+	default:
+		pr_err("invalid pgoff %lx for vma\n", curr->pgoff);
+		return -1;
+	}
+	return addr;
+}
+
+static bool find(const char *s1, const char *s2)
+{
+	if (s1 == NULL || s2 == NULL)
+		return NULL;
+
+	while (*s1 != '\0' && *s2 != '\0') {
+		if (*s1 == *s2) {
+			s1 += 1;
+			s2 += 1;
+		} else
+			s1 += 1;
+
+		if (*s2 == '\0')
+			return true;
+	}
+
+	return false;
+}
+
+static unsigned long distribute_restore_mapping(struct task_restore_args *args,
+						int i, int *step)
+{
+	VmaEntry *vma = args->vmas + i;
+	struct vma_names *vma_name = args->vma_names + i;
+
+	if (vma_entry_is(vma, VMA_AREA_CHR) && find(vma_name->name, HISI_SEC_DEV))
+		return restore_hisi_sec_mapping(args, i, step);
+	else
+		return restore_mapping(vma);
+}
+
 /*
  * This restores aio ring header, content, head and in-kernel position
  * of tail. To set tail, we write to /dev/null and use the fact this
@@ -1542,7 +1666,7 @@ int write_fork_pid(int pid)
 long __export_restore_task(struct task_restore_args *args)
 {
 	long ret = -1;
-	int i;
+	int i, step;
 	VmaEntry *vma_entry;
 	unsigned long va;
 	struct restore_vma_io *rio;
@@ -1691,7 +1815,7 @@ long __export_restore_task(struct task_restore_args *args)
 	/*
 	 * OK, lets try to map new one.
 	 */
-	for (i = 0; i < args->vmas_n; i++) {
+	for (i = 0, step = 1; i < args->vmas_n; i += step, step = 1) {
 		vma_entry = args->vmas + i;
 		vma_name = args->vma_names + i;
 
@@ -1708,7 +1832,7 @@ long __export_restore_task(struct task_restore_args *args)
 		if (vma_entry_is(vma_entry, VMA_PREMMAPED))
 			continue;
 
-		va = restore_mapping(vma_entry);
+		va = distribute_restore_mapping(args, i, &step);
 
 		if (va != vma_entry->start) {
 			pr_err("Can't restore %" PRIx64 " mapping with %lx\n", vma_entry->start, va);
diff --git a/criu/proc_parse.c b/criu/proc_parse.c
index 8913d93..daa54d9 100644
--- a/criu/proc_parse.c
+++ b/criu/proc_parse.c
@@ -41,6 +41,7 @@
 #include "path.h"
 #include "fault-injection.h"
 #include "memfd.h"
+#include "files-chr.h"
 
 #include "protobuf.h"
 #include "images/fdinfo.pb-c.h"
@@ -613,7 +614,8 @@ static int handle_vma(pid_t pid, struct vma_area *vma_area, const char *file_pat
 			/* NOTICE: if `--dump-char-dev` option is set, permmit
 			 * all char device memory area dumping.
 			 */
-			 if (strstr(file_path, "uverbs") != NULL) {
+			 if (strstr(file_path, "uverbs") != NULL
+				|| strstr(file_path, HISI_SEC_DEV) != NULL) {
 				int len = strlen(file_path) + 1;
 
 				vma_area->e->status |= VMA_AREA_CHR;
-- 
2.34.1