arch: Introduce user_fpregs_struct_t on ppc64

Signed-off-by: Pavel Emelyanov <xemul@virtuozzo.com>
[ldufour@linux.vnet.ibm.com: mostly rewrite to fix functional issue]
Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Signed-off-by: Pavel Emelyanov <xemul@virtuozzo.com>

criu/arch/ppc64/crtools.c

@@ -160,18 +160,16 @@ static UserPpc64FpstateEntry *copy_fp_regs(uint64_t *fpregs)
  * word registers are saved through FPR[0..32] and the remaining registers
  * are saved when saving the Altivec registers VR[0..32].
  */
-#define NVSXREG	32
 
-static UserPpc64FpstateEntry *get_fpu_regs(pid_t pid)
+static int get_fpu_regs(pid_t pid, user_fpregs_struct_t *fp)
 {
-	uint64_t fpregs[NFPREG];
-
-	if (ptrace(PTRACE_GETFPREGS, pid, 0, (void *)&fpregs) < 0) {
+	if (ptrace(PTRACE_GETFPREGS, pid, 0, (void *)&fp->fpregs) < 0) {
 		pr_perror("Couldn't get floating-point registers");
-		return NULL;;
+		return -1;
 	}
+	fp->flags |= USER_FPREGS_FL_FP;
 
-	return copy_fp_regs(fpregs);
+	return 0;
 }
 
 static void put_fpu_regs(mcontext_t *mc, UserPpc64FpstateEntry *fpe)
@@ -216,33 +214,22 @@ static UserPpc64VrstateEntry *copy_altivec_regs(__vector128 *vrregs)
 	return vse;
 }
 
-static UserPpc64VrstateEntry *get_altivec_regs(pid_t pid)
+static int get_altivec_regs(pid_t pid, user_fpregs_struct_t *fp)
 {
-	/* The kernel returns :
-	 *   32 Vector registers (128bit)
-	 *   VSCR (32bit) stored in a 128bit entry (odd)
-	 *   VRSAVE (32bit) store at the end.
-	 *
-	 * Kernel setup_sigcontext's comment mentions:
-	 * "Userland shall check AT_HWCAP to know whether it can rely on the
-	 * v_regs pointer or not"
-	 */
-	__vector128 vrregs[NVRREG];
-
-	if (ptrace(PTRACE_GETVRREGS, pid, 0, (void*)&vrregs) < 0) {
+	if (ptrace(PTRACE_GETVRREGS, pid, 0, (void*)&fp->vrregs) < 0) {
 		/* PTRACE_GETVRREGS returns EIO if Altivec is not supported.
 		 * This should not happen if msr_vec is set. */
 		if (errno != EIO) {
 			pr_perror("Couldn't get Altivec registers");
-			return (UserPpc64VrstateEntry*)-1L;
+			return -1;
 		}
 		pr_debug("Altivec not supported\n");
-		return NULL;
 	}
-
+	else {
 		pr_debug("Dumping Altivec registers\n");
-
-	return copy_altivec_regs(vrregs);
+		fp->flags |= USER_FPREGS_FL_ALTIVEC;
+	}
+	return 0;
 }
 
 static int put_altivec_regs(mcontext_t *mc, UserPpc64VrstateEntry *vse)
@@ -305,25 +292,24 @@ static UserPpc64VsxstateEntry* copy_vsx_regs(uint64_t *vsregs)
  * As a consequence, only the doubleword 1 of the 32 first VSX registers have
  * to be saved (the ones are returned by PTRACE_GETVSRREGS).
  */
-static UserPpc64VsxstateEntry *get_vsx_regs(pid_t pid)
+static int get_vsx_regs(pid_t pid, user_fpregs_struct_t *fp)
 {
-	uint64_t vsregs[NVSXREG];
-
-	if (ptrace(PTRACE_GETVSRREGS, pid, 0, (void*)&vsregs) < 0) {
+	if (ptrace(PTRACE_GETVSRREGS, pid, 0, (void*)fp->vsxregs) < 0) {
 		/*
 		 * EIO is returned in the case PTRACE_GETVRREGS is not
 		 * supported.
 		 */
-		if (errno == EIO) {
-			pr_debug("VSX register's dump not supported.\n");
-			return 0;
-		}
+		if (errno != EIO) {
 			pr_perror("Couldn't get VSX registers");
-		return (UserPpc64VsxstateEntry *)-1L;
+			return -1;
 		}
-
+		pr_debug("VSX register's dump not supported.\n");
+	}
+	else {
 		pr_debug("Dumping VSX registers\n");
-	return copy_vsx_regs(vsregs);
+		fp->flags |= USER_FPREGS_FL_VSX;
+	}
+	return 0;
 }
 
 static int put_vsx_regs(mcontext_t *mc, UserPpc64VsxstateEntry *vse)
@@ -434,29 +420,12 @@ static void xfree_tm_state(UserPpc64TmRegsEntry *tme)
 	}
 }
 
-static int get_tm_regs(pid_t pid, CoreEntry *core)
+static int get_tm_regs(pid_t pid, user_fpregs_struct_t *fpregs)
 {
-	struct {
-		uint64_t tfhar, texasr, tfiar;
-	} tm_spr_regs;
-	user_regs_struct_t regs;
-	uint64_t fpregs[NFPREG], vsxregs[32];
-	__vector128 vmxregs[NVRREG];
 	struct iovec iov;
-	UserPpc64TmRegsEntry *tme;
-	UserPpc64RegsEntry *gpregs = core->ti_ppc64->gpregs;
 
 	pr_debug("Dumping TM registers\n");
 
-	tme = xmalloc(sizeof(*tme));
-	if (!tme)
-		return -1;
-	user_ppc64_tm_regs_entry__init(tme);
-
-	tme->gpregs = allocate_gp_regs();
-	if (!tme->gpregs)
-		goto out_free;
-
 #define TM_REQUIRED	0
 #define TM_OPTIONAL	1
 #define PTRACE_GET_TM(s,n,c,u) do {					\
@@ -475,47 +444,29 @@ static int get_tm_regs(pid_t pid, CoreEntry *core)
 } while(0)
 
 	/* Get special registers */
-	PTRACE_GET_TM(tm_spr_regs, "SPR", NT_PPC_TM_SPR, TM_REQUIRED);
-	gpregs->has_tfhar 	= true;
-	gpregs->tfhar 		= tm_spr_regs.tfhar;
-	gpregs->has_texasr 	= true;
-	gpregs->texasr		= tm_spr_regs.texasr;
-	gpregs->has_tfiar 	= true;
-	gpregs->tfiar		= tm_spr_regs.tfiar;
+	PTRACE_GET_TM(fpregs->tm.tm_spr_regs, "SPR", NT_PPC_TM_SPR, TM_REQUIRED);
 
 	/* Get checkpointed regular registers */
-	PTRACE_GET_TM(regs, "GPR", NT_PPC_TM_CGPR, TM_REQUIRED);
-	copy_gp_regs(gpregs, &regs);
+	PTRACE_GET_TM(fpregs->tm.regs, "GPR", NT_PPC_TM_CGPR, TM_REQUIRED);
 
 	/* Get checkpointed FP registers */
-	PTRACE_GET_TM(fpregs, "FPR", NT_PPC_TM_CFPR, TM_OPTIONAL);
-	if (iov.iov_base) {
-		core->ti_ppc64->fpstate = copy_fp_regs(fpregs);
-		if (!core->ti_ppc64->fpstate)
-			goto out_free;
-	}
+	PTRACE_GET_TM(fpregs->tm.fpregs, "FPR", NT_PPC_TM_CFPR, TM_OPTIONAL);
+	if (iov.iov_base)
+		fpregs->tm.flags |= USER_FPREGS_FL_FP;
 
 	/* Get checkpointed VMX (Altivec) registers */
-	PTRACE_GET_TM(vmxregs, "VMX", NT_PPC_TM_CVMX, TM_OPTIONAL);
-	if (iov.iov_base) {
-		core->ti_ppc64->vrstate = copy_altivec_regs(vmxregs);
-		if (!core->ti_ppc64->vrstate)
-			goto out_free;
-	}
+	PTRACE_GET_TM(fpregs->tm.vrregs, "VMX", NT_PPC_TM_CVMX, TM_OPTIONAL);
+	if (iov.iov_base)
+		fpregs->tm.flags |= USER_FPREGS_FL_ALTIVEC;
 
 	/* Get checkpointed VSX registers */
-	PTRACE_GET_TM(vsxregs, "VSX", NT_PPC_TM_CVSX, TM_OPTIONAL);
-	if (iov.iov_base) {
-		core->ti_ppc64->vsxstate = copy_vsx_regs(vsxregs);
-		if (!core->ti_ppc64->vsxstate)
-			goto out_free;
-	}
+	PTRACE_GET_TM(fpregs->tm.vsxregs, "VSX", NT_PPC_TM_CVSX, TM_OPTIONAL);
+	if (iov.iov_base)
+		fpregs->tm.flags |= USER_FPREGS_FL_VSX;
 
-	core->ti_ppc64->tmstate = tme;
 	return 0;
 
 out_free:
-	xfree_tm_state(tme);
 	return -1;	/* still failing the checkpoint */
 }
 
@@ -553,13 +504,9 @@ static int put_tm_regs(struct rt_sigframe *f, UserPpc64TmRegsEntry *tme)
 }
 
 /****************************************************************************/
-int get_task_regs(pid_t pid, user_regs_struct_t regs, CoreEntry *core)
+static int __get_task_regs(pid_t pid, user_regs_struct_t *regs,
+			   user_fpregs_struct_t *fpregs)
 {
-	UserPpc64RegsEntry *gpregs;
-	UserPpc64FpstateEntry **fpstate;
-	UserPpc64VrstateEntry **vrstate;
-	UserPpc64VsxstateEntry **vsxstate;
-
 	pr_info("Dumping GP/FPU registers for %d\n", pid);
 
 	/*
@@ -570,59 +517,167 @@ int get_task_regs(pid_t pid, user_regs_struct_t regs, CoreEntry *core)
 #define TRAP(r)              ((r).trap & ~0xF)
 #endif
 
-	if (TRAP(regs) == 0x0C00 && regs.ccr & 0x10000000) {
+	if (TRAP(*regs) == 0x0C00 && regs->ccr & 0x10000000) {
 		/* Restart the system call */
-		switch (regs.gpr[3]) {
+		switch (regs->gpr[3]) {
 		case ERESTARTNOHAND:
 		case ERESTARTSYS:
 		case ERESTARTNOINTR:
-			regs.gpr[3] = regs.orig_gpr3;
-			regs.nip -= 4;
+			regs->gpr[3] = regs->orig_gpr3;
+			regs->nip -= 4;
 			break;
 		case ERESTART_RESTARTBLOCK:
-			regs.gpr[0] = __NR_restart_syscall;
-			regs.nip -= 4;
+			regs->gpr[0] = __NR_restart_syscall;
+			regs->nip -= 4;
 			break;
 		}
 	}
 
 	/* Resetting trap since we are now coming from user space. */
-	regs.trap = 0;
+	regs->trap = 0;
 
+	fpregs->flags = 0;
 	/*
 	 * Check for Transactional Memory operation in progress.
 	 * Until we have support of TM register's state through the ptrace API,
 	 * we can't checkpoint process with TM operation in progress (almost
 	 * impossible) or suspended (easy to get).
 	 */
-	if (MSR_TM_ACTIVE(regs.msr)) {
+	if (MSR_TM_ACTIVE(regs->msr)) {
 		pr_debug("Task %d has %s TM operation at 0x%lx\n",
 			 pid,
-			 (regs.msr & MSR_TMS) ? "a suspended" : "an active",
-			 regs.nip);
-		if (get_tm_regs(pid, core))
+			 (regs->msr & MSR_TMS) ? "a suspended" : "an active",
+			 regs->nip);
+		if (get_tm_regs(pid, fpregs))
+			return -1;
+		fpregs->flags = USER_FPREGS_FL_TM;
+	}
+
+	if (get_fpu_regs(pid, fpregs))
+		return -1;
+
+	if (get_altivec_regs(pid, fpregs))
+		return -1;
+
+	if (fpregs->flags & USER_FPREGS_FL_ALTIVEC) {
+		/*
+		 * Save the VSX registers if Altivec registers are supported
+		 */
+		if (get_vsx_regs(pid, fpregs))
+			return -1;
+	}
+	return 0;
+}
+
+static int copy_tm_regs(user_regs_struct_t *regs, user_fpregs_struct_t *fpregs,
+			CoreEntry *core)
+{
+	UserPpc64TmRegsEntry *tme;
+        UserPpc64RegsEntry *gpregs = core->ti_ppc64->gpregs;
+
+	pr_debug("Copying TM registers\n");
+	        tme = xmalloc(sizeof(*tme));
+        if (!tme)
                 return -1;
 
+	user_ppc64_tm_regs_entry__init(tme);
+
+        tme->gpregs = allocate_gp_regs();
+        if (!tme->gpregs)
+                goto out_free;
+
+        gpregs->has_tfhar       = true;
+        gpregs->tfhar           = fpregs->tm.tm_spr_regs.tfhar;
+        gpregs->has_texasr      = true;
+        gpregs->texasr          = fpregs->tm.tm_spr_regs.texasr;
+        gpregs->has_tfiar       = true;
+        gpregs->tfiar           = fpregs->tm.tm_spr_regs.tfiar;
+
+
+	/* This is the checkpointed state, we must save it in place of the
+	 * current state because the signal handler is made in this way.
+	 * We invert the 2 states instead of when building the signal frame,
+	 * because we can't modify the gpregs manipulated by the common layer.
+	 */
+	copy_gp_regs(gpregs, &fpregs->tm.regs);
+
+	if (fpregs->tm.flags & USER_FPREGS_FL_FP) {
+                core->ti_ppc64->fpstate = copy_fp_regs(fpregs->tm.fpregs);
+                if (!core->ti_ppc64->fpstate)
+                        goto out_free;
+        }
+
+	if (fpregs->tm.flags & USER_FPREGS_FL_ALTIVEC) {
+                core->ti_ppc64->vrstate = copy_altivec_regs(fpregs->tm.vrregs);
+                if (!core->ti_ppc64->vrstate)
+                        goto out_free;
+
+                /*
+                 * Force the MSR_VEC bit of the restored MSR otherwise the
+                 * kernel will not restore them from the signal frame.
+                 */
+                gpregs->msr |= MSR_VEC;
+
+		if (fpregs->tm.flags & USER_FPREGS_FL_VSX) {
+			core->ti_ppc64->vsxstate = copy_vsx_regs(fpregs->tm.vsxregs);
+			if (!core->ti_ppc64->vsxstate)
+				goto out_free;
+			/*
+                         * Force the MSR_VSX bit of the restored MSR otherwise
+                         * the kernel will not restore them from the signal
+                         * frame.
+                         */
+                        gpregs->msr |= MSR_VSX;
+		}
+	}
+
+	core->ti_ppc64->tmstate = tme;
+	return 0;
+
+out_free:
+	xfree_tm_state(tme);
+	return -1;
+}
+
+static int __copy_task_regs(user_regs_struct_t *regs,
+			    user_fpregs_struct_t *fpregs,
+			    CoreEntry *core)
+{
+	UserPpc64RegsEntry *gpregs;
+	UserPpc64FpstateEntry **fpstate;
+	UserPpc64VrstateEntry **vrstate;
+	UserPpc64VsxstateEntry **vsxstate;
+
+	/* Copy retrieved registers in the proto data
+	 * If TM is in the loop we switch the saved register set because
+	 * the signal frame is built with checkpointed registers on top to not
+	 * confused TM unaware process, while ptrace is retrieving the
+	 * checkpointed set through the TM specific ELF notes.
+	 */
+	if (fpregs->flags & USER_FPREGS_FL_TM) {
+		if (copy_tm_regs(regs, fpregs, core))
+			return -1;
 		gpregs = core->ti_ppc64->tmstate->gpregs;
 		fpstate = &(core->ti_ppc64->tmstate->fpstate);
 		vrstate = &(core->ti_ppc64->tmstate->vrstate);
 		vsxstate = &(core->ti_ppc64->tmstate->vsxstate);
-	} else {
+	}
+	else {
 		gpregs = core->ti_ppc64->gpregs;
 		fpstate = &(core->ti_ppc64->fpstate);
 		vrstate = &(core->ti_ppc64->vrstate);
 		vsxstate = &(core->ti_ppc64->vsxstate);
         }
 
-	copy_gp_regs(gpregs, &regs);
-
-	*fpstate = get_fpu_regs(pid);
+	copy_gp_regs(gpregs, regs);
+	if (fpregs->flags & USER_FPREGS_FL_FP) {
+		*fpstate = copy_fp_regs(fpregs->fpregs);
 		if (!*fpstate)
 			return -1;
-
-	*vrstate = get_altivec_regs(pid);
-	if (*vrstate) {
-		if (*vrstate == (UserPpc64VrstateEntry*)-1L)
+	}
+	if (fpregs->flags & USER_FPREGS_FL_ALTIVEC) {
+		*vrstate = copy_altivec_regs(fpregs->vrregs);
+		if (!*vrstate)
 			return -1;
                 /*
                  * Force the MSR_VEC bit of the restored MSR otherwise the
@@ -630,12 +685,9 @@ int get_task_regs(pid_t pid, user_regs_struct_t regs, CoreEntry *core)
                  */
                 gpregs->msr |= MSR_VEC;
 
-		/*
-		 * Save the VSX registers if Altivec registers are supported
-		 */
-		*vsxstate = get_vsx_regs(pid);
-		if (*vsxstate) {
-			if (*vsxstate == (UserPpc64VsxstateEntry *)-1L)
+		if (fpregs->flags & USER_FPREGS_FL_VSX) {
+			*vsxstate = copy_vsx_regs(fpregs->vsxregs);
+			if (!*vsxstate)
 				return -1;
                         /*
                          * Force the MSR_VSX bit of the restored MSR otherwise
@@ -645,10 +697,23 @@ int get_task_regs(pid_t pid, user_regs_struct_t regs, CoreEntry *core)
                         gpregs->msr |= MSR_VSX;
 		}
 	}
-
 	return 0;
 }
 
+
+/****************************************************************************/
+int get_task_regs(pid_t pid, user_regs_struct_t regs, CoreEntry *core)
+{
+	user_fpregs_struct_t fpregs;
+	int ret;
+
+	ret = __get_task_regs(pid, &regs, &fpregs);
+	if (ret)
+		return ret;
+
+	return __copy_task_regs(&regs, &fpregs, core);
+}
+
 int arch_alloc_thread_info(CoreEntry *core)
 {
 	ThreadInfoPpc64 *ti_ppc64;

criu/arch/ppc64/include/asm/types.h

@@ -61,6 +61,32 @@ typedef struct {
         unsigned long result;           /* Result of a system call */
 } user_regs_struct_t;
 
+#define NVSXREG	32
+
+#define USER_FPREGS_FL_FP	0x00001
+#define USER_FPREGS_FL_ALTIVEC	0x00002
+#define USER_FPREGS_FL_VSX	0x00004
+#define USER_FPREGS_FL_TM	0x00010
+
+typedef struct {
+	uint64_t fpregs[NFPREG];
+	__vector128 vrregs[NVRREG];
+	uint64_t vsxregs[NVSXREG];
+
+	int flags;
+	struct tm_regs {
+		int flags;
+		struct {
+			uint64_t tfhar, texasr, tfiar;
+		} tm_spr_regs;
+		user_regs_struct_t regs;
+		uint64_t fpregs[NFPREG];
+		__vector128 vrregs[NVRREG];
+		uint64_t vsxregs[NVSXREG];
+	} tm;
+} user_fpregs_struct_t;
+
+
 typedef UserPpc64RegsEntry UserRegsEntry;
 
 #define CORE_ENTRY__MARCH	CORE_ENTRY__MARCH__PPC64