Date: Wed, 9 Nov 2016 19:07:06 +0100
From: Radim =?utf-8?B?S3LEjW3DocWZ?= <rkrcmar@redhat.com>
To: linux-kernel@vger.kernel.org, kvm@vger.kernel.org
Cc: Paolo Bonzini <pbonzini@redhat.com>, Bandan Das <bsd@redhat.com>,
        Nadav Amit <nadav.amit@gmail.com>
Subject: [PATCH v4] KVM: x86: emulate FXSAVE and FXRSTOR
Message-ID: <20161109180706.GA8015@potion>
References: <20161108195419.4607-1-rkrcmar@redhat.com>
 <20161108195419.4607-5-rkrcmar@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <20161108195419.4607-5-rkrcmar@redhat.com>
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 5684
Lines: 193

Internal errors were reported on 16 bit fxsave and fxrstor with ipxe.
Old Intels don't have unrestricted_guest, so we have to emulate them.

The patch takes advantage of the hardware implementation.

AMD and Intel differ in saving and restoring other fields in first 32
bytes.  A test wrote 0xff to the fxsave area, 0 to upper bits of MCSXR
in the fxsave area, executed fxrstor, rewrote the fxsave area to 0xee,
and executed fxsave:

  Intel (Nehalem):
    7f 1f 7f 7f ff 00 ff 07 ff ff ff ff ff ff 00 00
    ff ff ff ff ff ff 00 00 ff ff 00 00 ff ff 00 00
  Intel (Haswell -- deprecated FPU CS and FPU DS):
    7f 1f 7f 7f ff 00 ff 07 ff ff ff ff 00 00 00 00
    ff ff ff ff 00 00 00 00 ff ff 00 00 ff ff 00 00
  AMD (Opteron 2300-series):
    7f 1f 7f 7f ff 00 ee ee ee ee ee ee ee ee ee ee
    ee ee ee ee ee ee ee ee ff ff 00 00 ff ff 02 00

fxsave/fxrstor will only be emulated on early Intels, so KVM can't do
much to improve the situation.

Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
---
 v4:
 - add support for CR4.OSFXSR [Paolo]
   I tested hardware implementation on AMD and Intel: neither of them
   save XMM 0-7, but both save and restore MCSXR without CR4.OSFXSR, so
   KVM behaves like that as well.
 v3:
 - remove fxsave64 and extra colons at the end of asm to make old GCCs
   happy  (fxsave64 could have been implemented using other nmemonics,
   but there is no point when it won't be used + removing it makes the
   code nicer.)
 v2:
 - throws #GP to the guest when reserved MXCSR are set [Nadav]
 - returns internal emulation error if an exception is hit during
   execution
 - preserves XMM 8-15
---
 arch/x86/kvm/emulate.c | 129 ++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 128 insertions(+), 1 deletion(-)

diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index 6af3cac6ec89..7d4f9b7f06ee 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -3883,6 +3883,131 @@ static int em_movsxd(struct x86_emulate_ctxt *ctxt)
 	return X86EMUL_CONTINUE;
 }
 
+static int check_fxsr(struct x86_emulate_ctxt *ctxt)
+{
+	u32 eax = 1, ebx, ecx = 0, edx;
+
+	ctxt->ops->get_cpuid(ctxt, &eax, &ebx, &ecx, &edx);
+	if (!(edx & FFL(FXSR)))
+		return emulate_ud(ctxt);
+
+	if (ctxt->ops->get_cr(ctxt, 0) & (X86_CR0_TS | X86_CR0_EM))
+		return emulate_nm(ctxt);
+
+	/*
+	 * Don't emulate a case that should never be hit, instead of working
+	 * around a lack of fxsave64/fxrstor64 on old compilers.
+	 */
+	if (ctxt->mode >= X86EMUL_MODE_PROT64)
+		return X86EMUL_UNHANDLEABLE;
+
+	return X86EMUL_CONTINUE;
+}
+
+/*
+ * FXSAVE and FXRSTOR have 4 different formats depending on execution mode,
+ *  1) 16 bit mode
+ *  2) 32 bit mode
+ *     - like (1), but FIP and FDP (foo) are only 16 bit.  At least Intel CPUs
+ *       preserve whole 32 bit values, though, so (1) and (2) are the same wrt.
+ *       save and restore
+ *  3) 64-bit mode with REX.W prefix
+ *     - like (2), but XMM 8-15 are being saved and restored
+ *  4) 64-bit mode without REX.W prefix
+ *     - like (3), but FIP and FDP are 64 bit
+ *
+ * Emulation uses (3) for (1) and (2) and preserves XMM 8-15 to reach the
+ * desired result.  (4) is not emulated.
+ *
+ * Note: Guest and host CPUID.(EAX=07H,ECX=0H):EBX[bit 13] (deprecate FPU CS
+ * and FPU DS) should match.
+ */
+static int em_fxsave(struct x86_emulate_ctxt *ctxt)
+{
+	struct fxregs_state fx_state;
+	size_t size;
+	int rc;
+
+	rc = check_fxsr(ctxt);
+	if (rc != X86EMUL_CONTINUE)
+		return rc;
+
+	ctxt->ops->get_fpu(ctxt);
+
+	rc = asm_safe("fxsave %[fx]", , [fx] "+m"(fx_state));
+
+	ctxt->ops->put_fpu(ctxt);
+
+	if (rc != X86EMUL_CONTINUE)
+		return rc;
+
+	if (ctxt->ops->get_cr(ctxt, 4) & X86_CR4_OSFXSR)
+		size = offsetof(struct fxregs_state, xmm_space[8 * 16/4]);
+	else
+		size = offsetof(struct fxregs_state, xmm_space[0]);
+
+	return segmented_write(ctxt, ctxt->memop.addr.mem, &fx_state, size);
+}
+
+static int fxrstor_fixup(struct x86_emulate_ctxt *ctxt,
+		struct fxregs_state *new)
+{
+	int rc = X86EMUL_CONTINUE;
+	struct fxregs_state old;
+
+	rc = asm_safe("fxsave %[fx]", , [fx] "+m"(old));
+	if (rc != X86EMUL_CONTINUE)
+		return rc;
+
+	/*
+	 * 64 bit host will restore XMM 8-15, which is not correct on non-64
+	 * bit guests.  Load the current values in order to preserve 64 bit
+	 * XMMs after fxrstor.
+	 */
+#ifdef CONFIG_X86_64
+	/* XXX: accessing XMM 8-15 very awkwardly */
+	memcpy(&new->xmm_space[8 * 16/4], &old.xmm_space[8 * 16/4], 8 * 16);
+#endif
+
+	/*
+	 * Hardware doesn't save and restore XMM 0-7 without CR4.OSFXSR, but
+	 * does save and restore MXCSR.
+	 */
+	if (!(ctxt->ops->get_cr(ctxt, 4) & X86_CR4_OSFXSR))
+		memcpy(new->xmm_space, old.xmm_space, 8 * 16);
+
+	return rc;
+}
+
+static int em_fxrstor(struct x86_emulate_ctxt *ctxt)
+{
+	struct fxregs_state fx_state;
+	int rc;
+
+	rc = check_fxsr(ctxt);
+	if (rc != X86EMUL_CONTINUE)
+		return rc;
+
+	rc = segmented_read(ctxt, ctxt->memop.addr.mem, &fx_state, 512);
+	if (rc != X86EMUL_CONTINUE)
+		return rc;
+
+	if (fx_state.mxcsr >> 16)
+		return emulate_gp(ctxt, 0);
+
+	ctxt->ops->get_fpu(ctxt);
+
+	if (ctxt->mode < X86EMUL_MODE_PROT64)
+		rc = fxrstor_fixup(ctxt, &fx_state);
+
+	if (rc == X86EMUL_CONTINUE)
+		rc = asm_safe("fxrstor %[fx]", : [fx] "m"(fx_state));
+
+	ctxt->ops->put_fpu(ctxt);
+
+	return rc;
+}
+
 static bool valid_cr(int nr)
 {
 	switch (nr) {
@@ -4235,7 +4360,9 @@ static const struct gprefix pfx_0f_ae_7 = {
 };
 
 static const struct group_dual group15 = { {
-	N, N, N, N, N, N, N, GP(0, &pfx_0f_ae_7),
+	I(ModRM | Aligned16, em_fxsave),
+	I(ModRM | Aligned16, em_fxrstor),
+	N, N, N, N, N, GP(0, &pfx_0f_ae_7),
 }, {
 	N, N, N, N, N, N, N, N,
 } };
-- 
2.10.2