Merge branch 'checkpoint-restore:criu-dev' into sk-inet-ip-transparent

.circleci/config.yml

@@ -2,7 +2,7 @@ version: 2.1
 jobs:
   test-local-gcc:
     machine:
-      image: ubuntu-2004:202010-01
+      image: default
     working_directory: ~/criu
     steps:
       - checkout
@@ -11,7 +11,7 @@ jobs:
           command: sudo -E make -C scripts/ci local
   test-local-clang:
     machine:
-      image: ubuntu-2004:202010-01
+      image: default
     working_directory: ~/criu
     steps:
       - checkout

.github/workflows/docker-test.yml

@@ -12,7 +12,7 @@ jobs:
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
-        os: [ubuntu-20.04]
+        os: [ubuntu-22.04]
     steps:
     - uses: actions/checkout@v2
     - name: Run Docker Test (${{ matrix.os }})

criu/Makefile.crtools

@@ -92,6 +92,7 @@ obj-y			+= servicefd.o
 obj-y			+= pie-util-vdso.o
 obj-y			+= vdso.o
 obj-y			+= timens.o
+obj-y			+= timer.o
 obj-$(CONFIG_HAS_LIBBPF)	+= bpfmap.o
 obj-$(CONFIG_COMPAT)	+= pie-util-vdso-elf32.o
 CFLAGS_pie-util-vdso-elf32.o	+= -DCONFIG_VDSO_32

criu/cr-dump.c

@@ -86,6 +86,7 @@
 #include "pidfd-store.h"
 #include "apparmor.h"
 #include "asm/dump.h"
+#include "timer.h"
 
 /*
  * Architectures can overwrite this function to restore register sets that
@@ -157,6 +158,11 @@ static int dump_sched_info(int pid, ThreadCoreEntry *tc)
 	tc->has_sched_policy = true;
 	tc->sched_policy = ret;
 
+	/* The reset-on-fork flag might be used in combination
+	 * with SCHED_FIFO or SCHED_RR to reset the scheduling
+	 * policy/priority in child processes.
+	 */
+	ret &= ~SCHED_RESET_ON_FORK;
 	if ((ret == SCHED_RR) || (ret == SCHED_FIFO)) {
 		ret = syscall(__NR_sched_getparam, pid, &sp);
 		if (ret < 0) {

criu/cr-restore.c

@@ -98,6 +98,7 @@
 #include "restore.h"
 
 #include "cr-errno.h"
+#include "timer.h"
 
 #ifndef arch_export_restore_thread
 #define arch_export_restore_thread __export_restore_thread
@@ -118,7 +119,6 @@ static int restore_task_with_children(void *);
 static int sigreturn_restore(pid_t pid, struct task_restore_args *ta, unsigned long alen, CoreEntry *core);
 static int prepare_restorer_blob(void);
 static int prepare_rlimits(int pid, struct task_restore_args *, CoreEntry *core);
-static int prepare_posix_timers(int pid, struct task_restore_args *ta, CoreEntry *core);
 static int prepare_signals(int pid, struct task_restore_args *, CoreEntry *core);
 
 /*
@@ -882,7 +882,6 @@ static int prepare_proc_misc(pid_t pid, TaskCoreEntry *tc, struct task_restore_a
 	return 0;
 }
 
-static int prepare_itimers(int pid, struct task_restore_args *args, CoreEntry *core);
 static int prepare_mm(pid_t pid, struct task_restore_args *args);
 
 static int restore_one_alive_task(int pid, CoreEntry *core)
@@ -2719,245 +2718,6 @@ static long restorer_get_vma_hint(struct list_head *tgt_vma_list, struct list_he
 	return -1;
 }
 
-static inline int timeval_valid(struct timeval *tv)
-{
-	return (tv->tv_sec >= 0) && ((unsigned long)tv->tv_usec < USEC_PER_SEC);
-}
-
-static inline int decode_itimer(char *n, ItimerEntry *ie, struct itimerval *val)
-{
-	if (ie->isec == 0 && ie->iusec == 0) {
-		memzero_p(val);
-		return 0;
-	}
-
-	val->it_interval.tv_sec = ie->isec;
-	val->it_interval.tv_usec = ie->iusec;
-
-	if (!timeval_valid(&val->it_interval)) {
-		pr_err("Invalid timer interval\n");
-		return -1;
-	}
-
-	if (ie->vsec == 0 && ie->vusec == 0) {
-		/*
-		 * Remaining time was too short. Set it to
-		 * interval to make the timer armed and work.
-		 */
-		val->it_value.tv_sec = ie->isec;
-		val->it_value.tv_usec = ie->iusec;
-	} else {
-		val->it_value.tv_sec = ie->vsec;
-		val->it_value.tv_usec = ie->vusec;
-	}
-
-	if (!timeval_valid(&val->it_value)) {
-		pr_err("Invalid timer value\n");
-		return -1;
-	}
-
-	pr_info("Restored %s timer to %ld.%ld -> %ld.%ld\n", n, val->it_value.tv_sec, val->it_value.tv_usec,
-		val->it_interval.tv_sec, val->it_interval.tv_usec);
-
-	return 0;
-}
-
-/*
- * Legacy itimers restore from CR_FD_ITIMERS
- */
-
-static int prepare_itimers_from_fd(int pid, struct task_restore_args *args)
-{
-	int ret = -1;
-	struct cr_img *img;
-	ItimerEntry *ie;
-
-	if (!deprecated_ok("Itimers"))
-		return -1;
-
-	img = open_image(CR_FD_ITIMERS, O_RSTR, pid);
-	if (!img)
-		return -1;
-
-	ret = pb_read_one(img, &ie, PB_ITIMER);
-	if (ret < 0)
-		goto out;
-	ret = decode_itimer("real", ie, &args->itimers[0]);
-	itimer_entry__free_unpacked(ie, NULL);
-	if (ret < 0)
-		goto out;
-
-	ret = pb_read_one(img, &ie, PB_ITIMER);
-	if (ret < 0)
-		goto out;
-	ret = decode_itimer("virt", ie, &args->itimers[1]);
-	itimer_entry__free_unpacked(ie, NULL);
-	if (ret < 0)
-		goto out;
-
-	ret = pb_read_one(img, &ie, PB_ITIMER);
-	if (ret < 0)
-		goto out;
-	ret = decode_itimer("prof", ie, &args->itimers[2]);
-	itimer_entry__free_unpacked(ie, NULL);
-	if (ret < 0)
-		goto out;
-out:
-	close_image(img);
-	return ret;
-}
-
-static int prepare_itimers(int pid, struct task_restore_args *args, CoreEntry *core)
-{
-	int ret = 0;
-	TaskTimersEntry *tte = core->tc->timers;
-
-	if (!tte)
-		return prepare_itimers_from_fd(pid, args);
-
-	ret |= decode_itimer("real", tte->real, &args->itimers[0]);
-	ret |= decode_itimer("virt", tte->virt, &args->itimers[1]);
-	ret |= decode_itimer("prof", tte->prof, &args->itimers[2]);
-
-	return ret;
-}
-
-static inline int timespec_valid(struct timespec *ts)
-{
-	return (ts->tv_sec >= 0) && ((unsigned long)ts->tv_nsec < NSEC_PER_SEC);
-}
-
-static inline int decode_posix_timer(PosixTimerEntry *pte, struct restore_posix_timer *pt)
-{
-	pt->val.it_interval.tv_sec = pte->isec;
-	pt->val.it_interval.tv_nsec = pte->insec;
-
-	if (!timespec_valid(&pt->val.it_interval)) {
-		pr_err("Invalid timer interval(posix)\n");
-		return -1;
-	}
-
-	if (pte->vsec == 0 && pte->vnsec == 0) {
-		/*
-		 * Remaining time was too short. Set it to
-		 * interval to make the timer armed and work.
-		 */
-		pt->val.it_value.tv_sec = pte->isec;
-		pt->val.it_value.tv_nsec = pte->insec;
-	} else {
-		pt->val.it_value.tv_sec = pte->vsec;
-		pt->val.it_value.tv_nsec = pte->vnsec;
-	}
-
-	if (!timespec_valid(&pt->val.it_value)) {
-		pr_err("Invalid timer value(posix)\n");
-		return -1;
-	}
-
-	pt->spt.it_id = pte->it_id;
-	pt->spt.clock_id = pte->clock_id;
-	pt->spt.si_signo = pte->si_signo;
-	pt->spt.it_sigev_notify = pte->it_sigev_notify;
-	pt->spt.sival_ptr = decode_pointer(pte->sival_ptr);
-	pt->spt.notify_thread_id = pte->notify_thread_id;
-	pt->overrun = pte->overrun;
-
-	return 0;
-}
-
-static int cmp_posix_timer_proc_id(const void *p1, const void *p2)
-{
-	return ((struct restore_posix_timer *)p1)->spt.it_id - ((struct restore_posix_timer *)p2)->spt.it_id;
-}
-
-static void sort_posix_timers(struct task_restore_args *ta)
-{
-	void *tmem;
-
-	/*
-	 * This is required for restorer's create_posix_timers(),
-	 * it will probe them one-by-one for the desired ID, since
-	 * kernel doesn't provide another API for timer creation
-	 * with given ID.
-	 */
-
-	if (ta->posix_timers_n > 0) {
-		tmem = rst_mem_remap_ptr((unsigned long)ta->posix_timers, RM_PRIVATE);
-		qsort(tmem, ta->posix_timers_n, sizeof(struct restore_posix_timer), cmp_posix_timer_proc_id);
-	}
-}
-
-/*
- * Legacy posix timers restoration from CR_FD_POSIX_TIMERS
- */
-
-static int prepare_posix_timers_from_fd(int pid, struct task_restore_args *ta)
-{
-	struct cr_img *img;
-	int ret = -1;
-	struct restore_posix_timer *t;
-
-	if (!deprecated_ok("Posix timers"))
-		return -1;
-
-	img = open_image(CR_FD_POSIX_TIMERS, O_RSTR, pid);
-	if (!img)
-		return -1;
-
-	ta->posix_timers_n = 0;
-	while (1) {
-		PosixTimerEntry *pte;
-
-		ret = pb_read_one_eof(img, &pte, PB_POSIX_TIMER);
-		if (ret <= 0)
-			break;
-
-		t = rst_mem_alloc(sizeof(struct restore_posix_timer), RM_PRIVATE);
-		if (!t)
-			break;
-
-		ret = decode_posix_timer(pte, t);
-		if (ret < 0)
-			break;
-
-		posix_timer_entry__free_unpacked(pte, NULL);
-		ta->posix_timers_n++;
-	}
-
-	close_image(img);
-	if (!ret)
-		sort_posix_timers(ta);
-
-	return ret;
-}
-
-static int prepare_posix_timers(int pid, struct task_restore_args *ta, CoreEntry *core)
-{
-	int i, ret = -1;
-	TaskTimersEntry *tte = core->tc->timers;
-	struct restore_posix_timer *t;
-
-	ta->posix_timers = (struct restore_posix_timer *)rst_mem_align_cpos(RM_PRIVATE);
-
-	if (!tte)
-		return prepare_posix_timers_from_fd(pid, ta);
-
-	ta->posix_timers_n = tte->n_posix;
-	for (i = 0; i < ta->posix_timers_n; i++) {
-		t = rst_mem_alloc(sizeof(struct restore_posix_timer), RM_PRIVATE);
-		if (!t)
-			goto out;
-
-		if (decode_posix_timer(tte->posix[i], t))
-			goto out;
-	}
-
-	ret = 0;
-	sort_posix_timers(ta);
-out:
-	return ret;
-}
-
 static int prepare_mm(pid_t pid, struct task_restore_args *args)
 {
 	int exe_fd, i, ret = -1;
@@ -3057,7 +2817,7 @@ static int validate_sched_parm(struct rst_sched_param *sp)
 	if ((sp->nice < -20) || (sp->nice > 19))
 		return 0;
 
-	switch (sp->policy) {
+	switch (sp->policy & ~SCHED_RESET_ON_FORK) {
 	case SCHED_RR:
 	case SCHED_FIFO:
 		return ((sp->prio > 0) && (sp->prio < 100));

criu/include/parasite-syscall.h

@@ -22,11 +22,6 @@ struct parasite_ctl;
 struct parasite_thread_ctl;
 
 extern int parasite_dump_sigacts_seized(struct parasite_ctl *ctl, struct pstree_item *);
-extern int parasite_dump_itimers_seized(struct parasite_ctl *ctl, struct pstree_item *);
-
-struct proc_posix_timers_stat;
-extern int parasite_dump_posix_timers_seized(struct proc_posix_timers_stat *proc_args, struct parasite_ctl *ctl,
-					     struct pstree_item *);
 
 extern int parasite_dump_misc_seized(struct parasite_ctl *ctl, struct parasite_dump_misc *misc);
 extern int parasite_dump_creds(struct parasite_ctl *ctl, CredsEntry *ce);

criu/include/timer.h

@@ -0,0 +1,17 @@
+#ifndef __CR_TIMER_H__
+#define __CR_TIMER_H__
+
+#include "images/core.pb-c.h"
+
+struct task_restore_args;
+struct pstree_item;
+struct parasite_ctl;
+struct proc_posix_timers_stat;
+
+extern int prepare_itimers(int pid, struct task_restore_args *args, CoreEntry *core);
+extern int prepare_posix_timers(int pid, struct task_restore_args *ta, CoreEntry *core);
+
+extern int parasite_dump_itimers_seized(struct parasite_ctl *ctl, struct pstree_item *item);
+extern int parasite_dump_posix_timers_seized(struct proc_posix_timers_stat *proc_args, struct parasite_ctl *ctl,
+					     struct pstree_item *item);
+#endif

criu/mem.c

@@ -1057,7 +1057,7 @@ static int premap_priv_vmas(struct pstree_item *t, struct vm_area_list *vmas, vo
 			do {
 				if (pr->pe->vaddr + pr->pe->nr_pages * PAGE_SIZE <= vma->e->start)
 					continue;
-				if (pr->pe->vaddr > vma->e->end)
+				if (pr->pe->vaddr >= vma->e->end)
 					vma->e->status |= VMA_NO_PROT_WRITE;
 				break;
 			} while (pr->advance(pr));