Add support for the extended FPU states on amd64, both for native

64bit and 32bit ABIs.  As a side-effect, it enables AVX on capable
CPUs.

In particular:

- Query the CPU support for XSAVE, list of the supported extensions
  and the required size of FPU save area. The hw.use_xsave tunable is
  provided for disabling XSAVE, and hw.xsave_mask may be used to
  select the enabled extensions.

- Remove the FPU save area from PCB and dynamically allocate the
  (run-time sized) user save area on the top of the kernel stack,
  right above the PCB. Reorganize the thread0 PCB initialization to
  postpone it after BSP is queried for save area size.

- The dumppcb, stoppcbs and susppcbs now do not carry the FPU state as
  well. FPU state is only useful for suspend, where it is saved in
  dynamically allocated suspfpusave area.

- Use XSAVE and XRSTOR to save/restore FPU state, if supported and
  enabled.

- Define new mcontext_t flag _MC_HASFPXSTATE, indicating that
  mcontext_t has a valid pointer to out-of-struct extended FPU
  state. Signal handlers are supplied with stack-allocated fpu
  state. The sigreturn(2) and setcontext(2) syscall honour the flag,
  allowing the signal handlers to inspect and manipilate extended
  state in the interrupted context.

- The getcontext(2) never returns extended state, since there is no
  place in the fixed-sized mcontext_t to place variable-sized save
  area. And, since mcontext_t is embedded into ucontext_t, makes it
  impossible to fix in a reasonable way.  Instead of extending
  getcontext(2) syscall, provide a sysarch(2) facility to query
  extended FPU state.

- Add ptrace(2) support for getting and setting extended state; while
  there, implement missed PT_I386_{GET,SET}XMMREGS for 32bit binaries.

- Change fpu_kern KPI to not expose struct fpu_kern_ctx layout to
  consumers, making it opaque. Internally, struct fpu_kern_ctx now
  contains a space for the extended state. Convert in-kernel consumers
  of fpu_kern KPI both on i386 and amd64.

First version of the support for AVX was submitted by Tim Bird
<tim.bird am sony com> on behalf of Sony. This version was written
from scratch.

Tested by:	pho (previous version), Yamagi Burmeister <lists yamagi org>
MFC after:	1 month

2 files changed, 98 insertions, 23 deletions

diff --git a/sys/amd64/ia32/ia32_reg.c b/sys/amd64/ia32/ia32_reg.c
index 279df9a..71eed5e 100644
--- a/sys/amd64/ia32/ia32_reg.c
+++ b/sys/amd64/ia32/ia32_reg.c
@@ -155,7 +155,7 @@ fill_fpregs32(struct thread *td, struct fpreg32 *regs)
 	sv_87 = (struct save87 *)regs;
 	penv_87 = &sv_87->sv_env;
 	fpugetregs(td);
-	sv_fpu = &td->td_pcb->pcb_user_save;
+	sv_fpu = get_pcb_user_save_td(td);
 	penv_xmm = &sv_fpu->sv_env;
 	
 	/* FPU control/status */
@@ -187,7 +187,7 @@ set_fpregs32(struct thread *td, struct fpreg32 *regs)
 {
 	struct save87 *sv_87 = (struct save87 *)regs;
 	struct env87 *penv_87 = &sv_87->sv_env;
-	struct savefpu *sv_fpu = &td->td_pcb->pcb_user_save;
+	struct savefpu *sv_fpu = get_pcb_user_save_td(td);
 	struct envxmm *penv_xmm = &sv_fpu->sv_env;
 	int i;
 
diff --git a/sys/amd64/ia32/ia32_signal.c b/sys/amd64/ia32/ia32_signal.c
index 2f41870..09ec7ab 100644
--- a/sys/amd64/ia32/ia32_signal.c
+++ b/sys/amd64/ia32/ia32_signal.c
@@ -71,6 +71,7 @@ __FBSDID("$FreeBSD$");
 #include <compat/freebsd32/freebsd32_signal.h>
 #include <compat/freebsd32/freebsd32_util.h>
 #include <compat/freebsd32/freebsd32_proto.h>
+#include <compat/freebsd32/freebsd32.h>
 #include <compat/ia32/ia32_signal.h>
 #include <machine/psl.h>
 #include <machine/segments.h>
@@ -83,15 +84,15 @@ __FBSDID("$FreeBSD$");
 #ifdef COMPAT_FREEBSD4
 static void freebsd4_ia32_sendsig(sig_t, ksiginfo_t *, sigset_t *);
 #endif
-static void ia32_get_fpcontext(struct thread *td, struct ia32_mcontext *mcp);
-static int ia32_set_fpcontext(struct thread *td, const struct ia32_mcontext *mcp);
 
 #define	CS_SECURE(cs)		(ISPL(cs) == SEL_UPL)
 #define	EFL_SECURE(ef, oef)	((((ef) ^ (oef)) & ~PSL_USERCHANGE) == 0)
 
 static void
-ia32_get_fpcontext(struct thread *td, struct ia32_mcontext *mcp)
+ia32_get_fpcontext(struct thread *td, struct ia32_mcontext *mcp,
+    char *xfpusave, size_t xfpusave_len)
 {
+	size_t max_len, len;
 
 	/*
 	 * XXX Format of 64bit and 32bit FXSAVE areas differs. FXSAVE
@@ -100,28 +101,43 @@ ia32_get_fpcontext(struct thread *td, struct ia32_mcontext *mcp)
 	 * for now, it should be irrelevant for most applications.
 	 */
 	mcp->mc_ownedfp = fpugetregs(td);
-	bcopy(&td->td_pcb->pcb_user_save, &mcp->mc_fpstate,
+	bcopy(get_pcb_user_save_td(td), &mcp->mc_fpstate,
 	    sizeof(mcp->mc_fpstate));
 	mcp->mc_fpformat = fpuformat();
+	if (!use_xsave || xfpusave_len == 0)
+		return;
+	max_len = cpu_max_ext_state_size - sizeof(struct savefpu);
+	len = xfpusave_len;
+	if (len > max_len) {
+		len = max_len;
+		bzero(xfpusave + max_len, len - max_len);
+	}
+	mcp->mc_flags |= _MC_HASFPXSTATE;
+	mcp->mc_xfpustate_len = len;
+	bcopy(get_pcb_user_save_td(td) + 1, xfpusave, len);
 }
 
 static int
-ia32_set_fpcontext(struct thread *td, const struct ia32_mcontext *mcp)
+ia32_set_fpcontext(struct thread *td, const struct ia32_mcontext *mcp,
+    char *xfpustate, size_t xfpustate_len)
 {
+	int error;
 
 	if (mcp->mc_fpformat == _MC_FPFMT_NODEV)
 		return (0);
 	else if (mcp->mc_fpformat != _MC_FPFMT_XMM)
 		return (EINVAL);
-	else if (mcp->mc_ownedfp == _MC_FPOWNED_NONE)
+	else if (mcp->mc_ownedfp == _MC_FPOWNED_NONE) {
 		/* We don't care what state is left in the FPU or PCB. */
 		fpstate_drop(td);
-	else if (mcp->mc_ownedfp == _MC_FPOWNED_FPU ||
+		error = 0;
+	} else if (mcp->mc_ownedfp == _MC_FPOWNED_FPU ||
 	    mcp->mc_ownedfp == _MC_FPOWNED_PCB) {
-		fpusetregs(td, (struct savefpu *)&mcp->mc_fpstate);
+		error = fpusetregs(td, (struct savefpu *)&mcp->mc_fpstate,
+		    xfpustate, xfpustate_len);
 	} else
 		return (EINVAL);
-	return (0);
+	return (error);
 }
 
 /*
@@ -164,10 +180,12 @@ ia32_get_mcontext(struct thread *td, struct ia32_mcontext *mcp, int flags)
 	mcp->mc_esp = tp->tf_rsp;
 	mcp->mc_ss = tp->tf_ss;
 	mcp->mc_len = sizeof(*mcp);
-	ia32_get_fpcontext(td, mcp);
+	mcp->mc_flags = tp->tf_flags;
+	ia32_get_fpcontext(td, mcp, NULL, 0);
 	mcp->mc_fsbase = pcb->pcb_fsbase;
 	mcp->mc_gsbase = pcb->pcb_gsbase;
-	bzero(mcp->mc_spare1, sizeof(mcp->mc_spare1));
+	mcp->mc_xfpustate = 0;
+	mcp->mc_xfpustate_len = 0;
 	bzero(mcp->mc_spare2, sizeof(mcp->mc_spare2));
 	set_pcb_flags(pcb, PCB_FULL_IRET);
 	return (0);
@@ -183,6 +201,7 @@ static int
 ia32_set_mcontext(struct thread *td, const struct ia32_mcontext *mcp)
 {
 	struct trapframe *tp;
+	char *xfpustate;
 	long rflags;
 	int ret;
 
@@ -191,7 +210,18 @@ ia32_set_mcontext(struct thread *td, const struct ia32_mcontext *mcp)
 		return (EINVAL);
 	rflags = (mcp->mc_eflags & PSL_USERCHANGE) |
 	    (tp->tf_rflags & ~PSL_USERCHANGE);
-	ret = ia32_set_fpcontext(td, mcp);
+	if (mcp->mc_flags & _MC_IA32_HASFPXSTATE) {
+		if (mcp->mc_xfpustate_len > cpu_max_ext_state_size -
+		    sizeof(struct savefpu))
+			return (EINVAL);
+		xfpustate = __builtin_alloca(mcp->mc_xfpustate_len);
+		ret = copyin(PTRIN(mcp->mc_xfpustate), xfpustate,
+		    mcp->mc_xfpustate_len);
+		if (ret != 0)
+			return (ret);
+	} else
+		xfpustate = NULL;
+	ret = ia32_set_fpcontext(td, mcp, xfpustate, mcp->mc_xfpustate_len);
 	if (ret != 0)
 		return (ret);
 	tp->tf_gs = mcp->mc_gs;
@@ -529,6 +559,8 @@ ia32_sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
 	struct sigacts *psp;
 	char *sp;
 	struct trapframe *regs;
+	char *xfpusave;
+	size_t xfpusave_len;
 	int oonstack;
 	int sig;
 
@@ -554,6 +586,14 @@ ia32_sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
 	regs = td->td_frame;
 	oonstack = sigonstack(regs->tf_rsp);
 
+	if (cpu_max_ext_state_size > sizeof(struct savefpu) && use_xsave) {
+		xfpusave_len = cpu_max_ext_state_size - sizeof(struct savefpu);
+		xfpusave = __builtin_alloca(xfpusave_len);
+	} else {
+		xfpusave_len = 0;
+		xfpusave = NULL;
+	}
+
 	/* Save user context. */
 	bzero(&sf, sizeof(sf));
 	sf.sf_uc.uc_sigmask = *mask;
@@ -582,7 +622,7 @@ ia32_sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
 	sf.sf_uc.uc_mcontext.mc_fs = regs->tf_fs;
 	sf.sf_uc.uc_mcontext.mc_gs = regs->tf_gs;
 	sf.sf_uc.uc_mcontext.mc_len = sizeof(sf.sf_uc.uc_mcontext); /* magic */
-	ia32_get_fpcontext(td, &sf.sf_uc.uc_mcontext);
+	ia32_get_fpcontext(td, &sf.sf_uc.uc_mcontext, xfpusave, xfpusave_len);
 	fpstate_drop(td);
 	sf.sf_uc.uc_mcontext.mc_fsbase = td->td_pcb->pcb_fsbase;
 	sf.sf_uc.uc_mcontext.mc_gsbase = td->td_pcb->pcb_gsbase;
@@ -590,11 +630,16 @@ ia32_sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
 
 	/* Allocate space for the signal handler context. */
 	if ((td->td_pflags & TDP_ALTSTACK) != 0 && !oonstack &&
-	    SIGISMEMBER(psp->ps_sigonstack, sig)) {
-		sp = td->td_sigstk.ss_sp +
-		    td->td_sigstk.ss_size - sizeof(sf);
-	} else
-		sp = (char *)regs->tf_rsp - sizeof(sf);
+	    SIGISMEMBER(psp->ps_sigonstack, sig))
+		sp = td->td_sigstk.ss_sp + td->td_sigstk.ss_size;
+	else
+		sp = (char *)regs->tf_rsp;
+	if (xfpusave != NULL) {
+		sp -= xfpusave_len;
+		sp = (char *)((unsigned long)sp & ~0x3Ful);
+		sf.sf_uc.uc_mcontext.mc_xfpustate = (register_t)sp;
+	}
+	sp -= sizeof(sf);
 	/* Align to 16 bytes. */
 	sfp = (struct ia32_sigframe *)((uintptr_t)sp & ~0xF);
 	PROC_UNLOCK(p);
@@ -626,7 +671,10 @@ ia32_sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
 	/*
 	 * Copy the sigframe out to the user's stack.
 	 */
-	if (copyout(&sf, sfp, sizeof(*sfp)) != 0) {
+	if (copyout(&sf, sfp, sizeof(*sfp)) != 0 ||
+	    (xfpusave != NULL && copyout(xfpusave,
+	    PTRIN(sf.sf_uc.uc_mcontext.mc_xfpustate), xfpusave_len)
+	    != 0)) {
 #ifdef DEBUG
 		printf("process %ld has trashed its stack\n", (long)p->p_pid);
 #endif
@@ -812,6 +860,8 @@ freebsd32_sigreturn(td, uap)
 	struct ia32_ucontext uc;
 	struct trapframe *regs;
 	struct ia32_ucontext *ucp;
+	char *xfpustate;
+	size_t xfpustate_len;
 	int cs, eflags, error, ret;
 	ksiginfo_t ksi;
 
@@ -858,9 +908,34 @@ freebsd32_sigreturn(td, uap)
 		return (EINVAL);
 	}
 
-	ret = ia32_set_fpcontext(td, &ucp->uc_mcontext);
-	if (ret != 0)
+	if ((ucp->uc_mcontext.mc_flags & _MC_HASFPXSTATE) != 0) {
+		xfpustate_len = uc.uc_mcontext.mc_xfpustate_len;
+		if (xfpustate_len > cpu_max_ext_state_size -
+		    sizeof(struct savefpu)) {
+			uprintf("pid %d (%s): sigreturn xfpusave_len = 0x%zx\n",
+			    td->td_proc->p_pid, td->td_name, xfpustate_len);
+			return (EINVAL);
+		}
+		xfpustate = __builtin_alloca(xfpustate_len);
+		error = copyin(PTRIN(ucp->uc_mcontext.mc_xfpustate),
+		    xfpustate, xfpustate_len);
+		if (error != 0) {
+			uprintf(
+	"pid %d (%s): sigreturn copying xfpustate failed\n",
+			    td->td_proc->p_pid, td->td_name);
+			return (error);
+		}
+	} else {
+		xfpustate = NULL;
+		xfpustate_len = 0;
+	}
+	ret = ia32_set_fpcontext(td, &ucp->uc_mcontext, xfpustate,
+	    xfpustate_len);
+	if (ret != 0) {
+		uprintf("pid %d (%s): sigreturn set_fpcontext err %d\n",
+		    td->td_proc->p_pid, td->td_name, ret);
 		return (ret);
+	}
 
 	regs->tf_rdi = ucp->uc_mcontext.mc_edi;
 	regs->tf_rsi = ucp->uc_mcontext.mc_esi;