amd64 pmap: microoptimize local shootdowns for PCID PTI configurations

When pmap operates in PTI mode, we must reload %cr3 on return to
userspace.  In non-PCID mode the reload always flushes all non-global
TLB entries and we take advantage of it by only invalidating the KPT
TLB entries (there is no cached UPT entries at all).

In PCID mode, we flush both KPT and UPT TLB explicitly, but we can
take advantage of the fact that PCID mode command to reload %cr3
includes a flag to flush/not flush target TLB.  In particular, we can
avoid the flush for UPT, instead record that load of pc_ucr3 into %cr3
on return to usermode should be flushing.  This is done by providing
either all-1s or ~CR3_PCID_MASK in pc_ucr3_load_mask.  The mask is
automatically reset to all-1s on return to usermode.

Similarly, we can avoid flushing UPT TLB on context switch, replacing
it by setting pc_ucr3_load_mask.  This unifies INVPCID and non-INVPCID
PTI ifunc, leaving only 4 cases instead of 6.  This trick is also
applicable both to the TLB shootdown IPI handlers, since handlers
interrupt the target thread.

But then we need to check pc_curpmap in handlers, and this would
reopen the same race for INVPCID machines as was fixed in r306350 for
non-INVPCID.  To not introduce the same bug, unconditionally do
spinlock_enter() in pmap_activate().

Reviewed by:	alc, markj
Tested by:	pho
Sponsored by:	The FreeBSD Foundation
MFC after:	3 weeks
Differential revision:	https://reviews.freebsd.org/D25483

7 files changed, 84 insertions, 110 deletions

diff --git a/sys/amd64/amd64/exception.S b/sys/amd64/amd64/exception.S
index a005c3b70591..b85471b93169 100644
--- a/sys/amd64/amd64/exception.S
+++ b/sys/amd64/amd64/exception.S
@@ -47,6 +47,7 @@
 #include <machine/asmacros.h>
 #include <machine/trap.h>
 #include <machine/specialreg.h>
+#include <machine/pmap.h>
 
 #ifdef KDTRACE_HOOKS
 	.bss
@@ -607,8 +608,10 @@ fast_syscall_common:
 	cmpq	$~0,PCPU(UCR3)
 	je	2f
 	movq	PCPU(UCR3),%r9
+	andq	PCPU(UCR3_LOAD_MASK),%r9
 	movq	%r9,%cr3
 2:	xorl	%r9d,%r9d
+	movq	$PMAP_UCR3_NOMASK,PCPU(UCR3_LOAD_MASK)
 	swapgs
 	sysretq
 
@@ -1262,6 +1265,8 @@ ld_regs:
 	movq	TF_SS(%rsp),%rax
 	movq	%rax,PTI_SS(%rdx)
 	movq	PCPU(UCR3),%rax
+	andq	PCPU(UCR3_LOAD_MASK),%rax
+	movq	$PMAP_UCR3_NOMASK,PCPU(UCR3_LOAD_MASK)
 	swapgs
 	movq	%rdx,%rsp
 	movq	%rax,%cr3
diff --git a/sys/amd64/amd64/genassym.c b/sys/amd64/amd64/genassym.c
index 78b18001e97b..ec3707ce41f9 100644
--- a/sys/amd64/amd64/genassym.c
+++ b/sys/amd64/amd64/genassym.c
@@ -230,6 +230,7 @@ ASSYM(PC_TSS, offsetof(struct pcpu, pc_tss));
 ASSYM(PC_PM_SAVE_CNT, offsetof(struct pcpu, pc_pm_save_cnt));
 ASSYM(PC_KCR3, offsetof(struct pcpu, pc_kcr3));
 ASSYM(PC_UCR3, offsetof(struct pcpu, pc_ucr3));
+ASSYM(PC_UCR3_LOAD_MASK, offsetof(struct pcpu, pc_ucr3_load_mask));
 ASSYM(PC_SAVED_UCR3, offsetof(struct pcpu, pc_saved_ucr3));
 ASSYM(PC_PTI_STACK, offsetof(struct pcpu, pc_pti_stack));
 ASSYM(PC_PTI_STACK_SZ, PC_PTI_STACK_SZ);
diff --git a/sys/amd64/amd64/machdep.c b/sys/amd64/amd64/machdep.c
index 1a07080c5daf..294cd8100d3b 100644
--- a/sys/amd64/amd64/machdep.c
+++ b/sys/amd64/amd64/machdep.c
@@ -1562,6 +1562,7 @@ amd64_bsp_pcpu_init1(struct pcpu *pc)
 	PCPU_SET(ldt, (struct system_segment_descriptor *)&gdt[GUSERLDT_SEL]);
 	PCPU_SET(fs32p, &gdt[GUFS32_SEL]);
 	PCPU_SET(gs32p, &gdt[GUGS32_SEL]);
+	PCPU_SET(ucr3_load_mask, PMAP_UCR3_NOMASK);
 	PCPU_SET(smp_tlb_gen, 1);
 }
 
diff --git a/sys/amd64/amd64/mp_machdep.c b/sys/amd64/amd64/mp_machdep.c
index 4d2df92664c7..ca95013d9c0b 100644
--- a/sys/amd64/amd64/mp_machdep.c
+++ b/sys/amd64/amd64/mp_machdep.c
@@ -283,6 +283,7 @@ init_secondary(void)
 	pc->pc_fs32p = &gdt[GUFS32_SEL];
 	pc->pc_gs32p = &gdt[GUGS32_SEL];
 	pc->pc_ldt = (struct system_segment_descriptor *)&gdt[GUSERLDT_SEL];
+	pc->pc_ucr3_load_mask = PMAP_UCR3_NOMASK;
 	/* See comment in pmap_bootstrap(). */
 	pc->pc_pcid_next = PMAP_PCID_KERN + 2;
 	pc->pc_pcid_gen = 1;
@@ -821,15 +822,14 @@ invltlb_invpcid_pti_handler(pmap_t smp_tlb_pmap)
 		invpcid(&d, INVPCID_CTXGLOB);
 	} else {
 		invpcid(&d, INVPCID_CTX);
-		d.pcid |= PMAP_PCID_USER_PT;
-		invpcid(&d, INVPCID_CTX);
+		if (smp_tlb_pmap == PCPU_GET(curpmap))
+			PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE);
 	}
 }
 
 static void
 invltlb_pcid_handler(pmap_t smp_tlb_pmap)
 {
-	uint64_t kcr3, ucr3;
 	uint32_t pcid;
   
 #ifdef COUNT_XINVLTLB_HITS
@@ -849,15 +849,11 @@ invltlb_pcid_handler(pmap_t smp_tlb_pmap)
 		 * invalidation when switching to the pmap on this
 		 * CPU.
 		 */
-		if (PCPU_GET(curpmap) == smp_tlb_pmap) {
+		if (smp_tlb_pmap == PCPU_GET(curpmap)) {
 			pcid = smp_tlb_pmap->pm_pcids[PCPU_GET(cpuid)].pm_pcid;
-			kcr3 = smp_tlb_pmap->pm_cr3 | pcid;
-			ucr3 = smp_tlb_pmap->pm_ucr3;
-			if (ucr3 != PMAP_NO_CR3) {
-				ucr3 |= PMAP_PCID_USER_PT | pcid;
-				pmap_pti_pcid_invalidate(ucr3, kcr3);
-			} else
-				load_cr3(kcr3);
+			load_cr3(smp_tlb_pmap->pm_cr3 | pcid);
+			if (smp_tlb_pmap->pm_ucr3 != PMAP_NO_CR3)
+				PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE);
 		}
 	}
 }
@@ -888,7 +884,9 @@ invlpg_invpcid_handler(pmap_t smp_tlb_pmap, vm_offset_t smp_tlb_addr1)
 #endif /* COUNT_IPIS */
 
 	invlpg(smp_tlb_addr1);
-	if (smp_tlb_pmap->pm_ucr3 != PMAP_NO_CR3) {
+	if (smp_tlb_pmap == PCPU_GET(curpmap) &&
+	    smp_tlb_pmap->pm_ucr3 != PMAP_NO_CR3 &&
+	    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
 		d.pcid = smp_tlb_pmap->pm_pcids[PCPU_GET(cpuid)].pm_pcid |
 		    PMAP_PCID_USER_PT;
 		d.pad = 0;
@@ -912,7 +910,8 @@ invlpg_pcid_handler(pmap_t smp_tlb_pmap, vm_offset_t smp_tlb_addr1)
 
 	invlpg(smp_tlb_addr1);
 	if (smp_tlb_pmap == PCPU_GET(curpmap) &&
-	    (ucr3 = smp_tlb_pmap->pm_ucr3) != PMAP_NO_CR3) {
+	    (ucr3 = smp_tlb_pmap->pm_ucr3) != PMAP_NO_CR3 &&
+	    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
 		pcid = smp_tlb_pmap->pm_pcids[PCPU_GET(cpuid)].pm_pcid;
 		kcr3 = smp_tlb_pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
 		ucr3 |= pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE;
@@ -960,7 +959,9 @@ invlrng_invpcid_handler(pmap_t smp_tlb_pmap, vm_offset_t smp_tlb_addr1,
 		invlpg(addr);
 		addr += PAGE_SIZE;
 	} while (addr < addr2);
-	if (smp_tlb_pmap->pm_ucr3 != PMAP_NO_CR3) {
+	if (smp_tlb_pmap == PCPU_GET(curpmap) &&
+	    smp_tlb_pmap->pm_ucr3 != PMAP_NO_CR3 &&
+	    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
 		d.pcid = smp_tlb_pmap->pm_pcids[PCPU_GET(cpuid)].pm_pcid |
 		    PMAP_PCID_USER_PT;
 		d.pad = 0;
@@ -994,7 +995,8 @@ invlrng_pcid_handler(pmap_t smp_tlb_pmap, vm_offset_t smp_tlb_addr1,
 		addr += PAGE_SIZE;
 	} while (addr < addr2);
 	if (smp_tlb_pmap == PCPU_GET(curpmap) &&
-	    (ucr3 = smp_tlb_pmap->pm_ucr3) != PMAP_NO_CR3) {
+	    (ucr3 = smp_tlb_pmap->pm_ucr3) != PMAP_NO_CR3 &&
+	    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
 		pcid = smp_tlb_pmap->pm_pcids[PCPU_GET(cpuid)].pm_pcid;
 		kcr3 = smp_tlb_pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
 		ucr3 |= pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE;
diff --git a/sys/amd64/amd64/pmap.c b/sys/amd64/amd64/pmap.c
index fdcbff0a3dad..314c8df5b328 100644
--- a/sys/amd64/amd64/pmap.c
+++ b/sys/amd64/amd64/pmap.c
@@ -2520,7 +2520,16 @@ pmap_invalidate_page_pcid(pmap_t pmap, vm_offset_t va,
 
 	cpuid = PCPU_GET(cpuid);
 	if (pmap == PCPU_GET(curpmap)) {
-		if (pmap->pm_ucr3 != PMAP_NO_CR3) {
+		if (pmap->pm_ucr3 != PMAP_NO_CR3 &&
+		    /*
+		     * If we context-switched right after
+		     * PCPU_GET(ucr3_load_mask), we could read the
+		     * ~CR3_PCID_SAVE mask, which causes us to skip
+		     * the code below to invalidate user pages.  This
+		     * is handled in pmap_activate_sw_pcid_pti() by
+		     * clearing pm_gen if ucr3_load_mask is ~CR3_PCID_SAVE.
+		     */
+		    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
 			/*
 			 * Because pm_pcid is recalculated on a
 			 * context switch, we must disable switching.
@@ -2635,7 +2644,8 @@ pmap_invalidate_range_pcid(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
 
 	cpuid = PCPU_GET(cpuid);
 	if (pmap == PCPU_GET(curpmap)) {
-		if (pmap->pm_ucr3 != PMAP_NO_CR3) {
+		if (pmap->pm_ucr3 != PMAP_NO_CR3 &&
+		    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
 			critical_enter();
 			pcid = pmap->pm_pcids[cpuid].pm_pcid;
 			if (invpcid_works1) {
@@ -2736,7 +2746,7 @@ static inline void
 pmap_invalidate_all_pcid(pmap_t pmap, bool invpcid_works1)
 {
 	struct invpcid_descr d;
-	uint64_t kcr3, ucr3;
+	uint64_t kcr3;
 	uint32_t pcid;
 	u_int cpuid, i;
 
@@ -2757,20 +2767,12 @@ pmap_invalidate_all_pcid(pmap_t pmap, bool invpcid_works1)
 				d.pad = 0;
 				d.addr = 0;
 				invpcid(&d, INVPCID_CTX);
-				if (pmap->pm_ucr3 != PMAP_NO_CR3) {
-					d.pcid |= PMAP_PCID_USER_PT;
-					invpcid(&d, INVPCID_CTX);
-				}
 			} else {
 				kcr3 = pmap->pm_cr3 | pcid;
-				ucr3 = pmap->pm_ucr3;
-				if (ucr3 != PMAP_NO_CR3) {
-					ucr3 |= pcid | PMAP_PCID_USER_PT;
-					pmap_pti_pcid_invalidate(ucr3, kcr3);
-				} else {
-					load_cr3(kcr3);
-				}
+				load_cr3(kcr3);
 			}
+			if (pmap->pm_ucr3 != PMAP_NO_CR3)
+				PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE);
 			critical_exit();
 		} else
 			pmap->pm_pcids[cpuid].pm_gen = 0;
@@ -8816,12 +8818,23 @@ pmap_activate_sw_pti_post(struct thread *td, pmap_t pmap)
 	    PCPU_GET(pti_rsp0) : (uintptr_t)td->td_md.md_stack_base;
 }
 
-static void inline
-pmap_activate_sw_pcid_pti(pmap_t pmap, u_int cpuid, const bool invpcid_works1)
+static void
+pmap_activate_sw_pcid_pti(struct thread *td, pmap_t pmap, u_int cpuid)
 {
-	struct invpcid_descr d;
+	pmap_t old_pmap;
 	uint64_t cached, cr3, kcr3, ucr3;
 
+	KASSERT((read_rflags() & PSL_I) == 0,
+	    ("PCID needs interrupts disabled in pmap_activate_sw()"));
+
+	/* See the comment in pmap_invalidate_page_pcid(). */
+	if (PCPU_GET(ucr3_load_mask) != PMAP_UCR3_NOMASK) {
+		PCPU_SET(ucr3_load_mask, PMAP_UCR3_NOMASK);
+		old_pmap = PCPU_GET(curpmap);
+		MPASS(old_pmap->pm_ucr3 != PMAP_NO_CR3);
+		old_pmap->pm_pcids[cpuid].pm_gen = 0;
+	}
+
 	cached = pmap_pcid_alloc_checked(pmap, cpuid);
 	cr3 = rcr3();
 	if ((cr3 & ~CR3_PCID_MASK) != pmap->pm_cr3)
@@ -8831,68 +8844,14 @@ pmap_activate_sw_pcid_pti(pmap_t pmap, u_int cpuid, const bool invpcid_works1)
 	ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[cpuid].pm_pcid |
 	    PMAP_PCID_USER_PT;
 
-	if (!cached && pmap->pm_ucr3 != PMAP_NO_CR3) {
-		/*
-		 * Explicitly invalidate translations cached from the
-		 * user page table.  They are not automatically
-		 * flushed by reload of cr3 with the kernel page table
-		 * pointer above.
-		 *
-		 * Note that the if() condition is resolved statically
-		 * by using the function argument instead of
-		 * runtime-evaluated invpcid_works value.
-		 */
-		if (invpcid_works1) {
-			d.pcid = PMAP_PCID_USER_PT |
-			    pmap->pm_pcids[cpuid].pm_pcid;
-			d.pad = 0;
-			d.addr = 0;
-			invpcid(&d, INVPCID_CTX);
-		} else {
-			pmap_pti_pcid_invalidate(ucr3, kcr3);
-		}
-	}
+	if (!cached && pmap->pm_ucr3 != PMAP_NO_CR3)
+		PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE);
 
 	PCPU_SET(kcr3, kcr3 | CR3_PCID_SAVE);
 	PCPU_SET(ucr3, ucr3 | CR3_PCID_SAVE);
 	if (cached)
 		PCPU_INC(pm_save_cnt);
-}
 
-static void
-pmap_activate_sw_pcid_invpcid_pti(struct thread *td, pmap_t pmap, u_int cpuid)
-{
-
-	pmap_activate_sw_pcid_pti(pmap, cpuid, true);
-	pmap_activate_sw_pti_post(td, pmap);
-}
-
-static void
-pmap_activate_sw_pcid_noinvpcid_pti(struct thread *td, pmap_t pmap,
-    u_int cpuid)
-{
-	register_t rflags;
-
-	/*
-	 * If the INVPCID instruction is not available,
-	 * invltlb_pcid_handler() is used to handle an invalidate_all
-	 * IPI, which checks for curpmap == smp_tlb_pmap.  The below
-	 * sequence of operations has a window where %CR3 is loaded
-	 * with the new pmap's PML4 address, but the curpmap value has
-	 * not yet been updated.  This causes the invltlb IPI handler,
-	 * which is called between the updates, to execute as a NOP,
-	 * which leaves stale TLB entries.
-	 *
-	 * Note that the most typical use of pmap_activate_sw(), from
-	 * the context switch, is immune to this race, because
-	 * interrupts are disabled (while the thread lock is owned),
-	 * and the IPI happens after curpmap is updated.  Protect
-	 * other callers in a similar way, by disabling interrupts
-	 * around the %cr3 register reload and curpmap assignment.
-	 */
-	rflags = intr_disable();
-	pmap_activate_sw_pcid_pti(pmap, cpuid, false);
-	intr_restore(rflags);
 	pmap_activate_sw_pti_post(td, pmap);
 }
 
@@ -8902,6 +8861,9 @@ pmap_activate_sw_pcid_nopti(struct thread *td __unused, pmap_t pmap,
 {
 	uint64_t cached, cr3;
 
+	KASSERT((read_rflags() & PSL_I) == 0,
+	    ("PCID needs interrupts disabled in pmap_activate_sw()"));
+
 	cached = pmap_pcid_alloc_checked(pmap, cpuid);
 	cr3 = rcr3();
 	if (!cached || (cr3 & ~CR3_PCID_MASK) != pmap->pm_cr3)
@@ -8913,17 +8875,6 @@ pmap_activate_sw_pcid_nopti(struct thread *td __unused, pmap_t pmap,
 }
 
 static void
-pmap_activate_sw_pcid_noinvpcid_nopti(struct thread *td __unused, pmap_t pmap,
-    u_int cpuid)
-{
-	register_t rflags;
-
-	rflags = intr_disable();
-	pmap_activate_sw_pcid_nopti(td, pmap, cpuid);
-	intr_restore(rflags);
-}
-
-static void
 pmap_activate_sw_nopcid_nopti(struct thread *td __unused, pmap_t pmap,
     u_int cpuid __unused)
 {
@@ -8947,14 +8898,10 @@ DEFINE_IFUNC(static, void, pmap_activate_sw_mode, (struct thread *, pmap_t,
     u_int))
 {
 
-	if (pmap_pcid_enabled && pti && invpcid_works)
-		return (pmap_activate_sw_pcid_invpcid_pti);
-	else if (pmap_pcid_enabled && pti && !invpcid_works)
-		return (pmap_activate_sw_pcid_noinvpcid_pti);
-	else if (pmap_pcid_enabled && !pti && invpcid_works)
+	if (pmap_pcid_enabled && pti)
+		return (pmap_activate_sw_pcid_pti);
+	else if (pmap_pcid_enabled && !pti)
 		return (pmap_activate_sw_pcid_nopti);
-	else if (pmap_pcid_enabled && !pti && !invpcid_works)
-		return (pmap_activate_sw_pcid_noinvpcid_nopti);
 	else if (!pmap_pcid_enabled && pti)
 		return (pmap_activate_sw_nopcid_pti);
 	else /* if (!pmap_pcid_enabled && !pti) */
@@ -8991,10 +8938,26 @@ pmap_activate_sw(struct thread *td)
 void
 pmap_activate(struct thread *td)
 {
-
-	critical_enter();
+	/*
+	 * invltlb_{invpcid,}_pcid_handler() is used to handle an
+	 * invalidate_all IPI, which checks for curpmap ==
+	 * smp_tlb_pmap.  The below sequence of operations has a
+	 * window where %CR3 is loaded with the new pmap's PML4
+	 * address, but the curpmap value has not yet been updated.
+	 * This causes the invltlb IPI handler, which is called
+	 * between the updates, to execute as a NOP, which leaves
+	 * stale TLB entries.
+	 *
+	 * Note that the most common use of pmap_activate_sw(), from
+	 * a context switch, is immune to this race, because
+	 * interrupts are disabled (while the thread lock is owned),
+	 * so the IPI is delayed until after curpmap is updated.  Protect
+	 * other callers in a similar way, by disabling interrupts
+	 * around the %cr3 register reload and curpmap assignment.
+	 */
+	spinlock_enter();
 	pmap_activate_sw(td);
-	critical_exit();
+	spinlock_exit();
 }
 
 void
diff --git a/sys/amd64/include/pcpu.h b/sys/amd64/include/pcpu.h
index 22c6ed40aa20..4b163636c775 100644
--- a/sys/amd64/include/pcpu.h
+++ b/sys/amd64/include/pcpu.h
@@ -99,7 +99,8 @@ _Static_assert(sizeof(struct monitorbuf) == 128, "2x cache line");
 	uint64_t pc_smp_tlb_addr2;					\
 	uint32_t pc_smp_tlb_gen;					\
 	u_int	pc_smp_tlb_op;						\
-	char	__pad[2924]		/* pad to UMA_PCPU_ALLOC_SIZE */
+	uint64_t pc_ucr3_load_mask;					\
+	char	__pad[2916]		/* pad to UMA_PCPU_ALLOC_SIZE */
 
 #define	PC_DBREG_CMD_NONE	0
 #define	PC_DBREG_CMD_LOAD	1
diff --git a/sys/amd64/include/pmap.h b/sys/amd64/include/pmap.h
index ffa21a674583..e2d7a714511b 100644
--- a/sys/amd64/include/pmap.h
+++ b/sys/amd64/include/pmap.h
@@ -241,6 +241,7 @@
 #define	PMAP_PCID_USER_PT	0x800
 
 #define	PMAP_NO_CR3		(~0UL)
+#define	PMAP_UCR3_NOMASK	(~0UL)
 
 #ifndef LOCORE