Re: [PATCH v15 27/30] LoongArch: KVM: Implement vcpu world switch
From: zhaotianrui
Date: Wed Jun 28 2023 - 04:37:36 EST
在 2023/6/28 上午11:42, Jinyang He 写道:
On 2023-06-26 16:47, Tianrui Zhao wrote:
Implement LoongArch vcpu world switch, including vcpu enter guest and
vcpu exit from guest, both operations need to save or restore the host
and guest registers.
Reviewed-by: Bibo Mao <maobibo@xxxxxxxxxxx>
Signed-off-by: Tianrui Zhao <zhaotianrui@xxxxxxxxxxx>
---
arch/loongarch/kernel/asm-offsets.c | 32 +++
arch/loongarch/kvm/switch.S | 299 ++++++++++++++++++++++++++++
2 files changed, 331 insertions(+)
create mode 100644 arch/loongarch/kvm/switch.S
diff --git a/arch/loongarch/kernel/asm-offsets.c
b/arch/loongarch/kernel/asm-offsets.c
index 4bdb203fc66e..cb6c5a5afea9 100644
--- a/arch/loongarch/kernel/asm-offsets.c
+++ b/arch/loongarch/kernel/asm-offsets.c
@@ -9,6 +9,7 @@
#include <linux/mm.h>
#include <linux/kbuild.h>
#include <linux/suspend.h>
+#include <linux/kvm_host.h>
#include <asm/cpu-info.h>
#include <asm/ptrace.h>
#include <asm/processor.h>
@@ -272,3 +273,34 @@ void output_pbe_defines(void)
BLANK();
}
#endif
+
+static void __used output_kvm_defines(void)
+{
+ COMMENT(" KVM/LOONGARCH Specific offsets. ");
+
+ OFFSET(VCPU_FCSR0, kvm_vcpu_arch, fpu.fcsr);
+ OFFSET(VCPU_FCC, kvm_vcpu_arch, fpu.fcc);
+ BLANK();
+
+ OFFSET(KVM_VCPU_ARCH, kvm_vcpu, arch);
+ OFFSET(KVM_VCPU_KVM, kvm_vcpu, kvm);
+ OFFSET(KVM_VCPU_RUN, kvm_vcpu, run);
+ BLANK();
+
+ OFFSET(KVM_ARCH_HSP, kvm_vcpu_arch, host_sp);
+ OFFSET(KVM_ARCH_HTP, kvm_vcpu_arch, host_tp);
+ OFFSET(KVM_ARCH_HANDLE_EXIT, kvm_vcpu_arch, handle_exit);
+ OFFSET(KVM_ARCH_HPGD, kvm_vcpu_arch, host_pgd);
+ OFFSET(KVM_ARCH_GEENTRY, kvm_vcpu_arch, guest_eentry);
+ OFFSET(KVM_ARCH_GPC, kvm_vcpu_arch, pc);
+ OFFSET(KVM_ARCH_GGPR, kvm_vcpu_arch, gprs);
+ OFFSET(KVM_ARCH_HESTAT, kvm_vcpu_arch, host_estat);
+ OFFSET(KVM_ARCH_HBADV, kvm_vcpu_arch, badv);
+ OFFSET(KVM_ARCH_HBADI, kvm_vcpu_arch, badi);
+ OFFSET(KVM_ARCH_HECFG, kvm_vcpu_arch, host_ecfg);
+ OFFSET(KVM_ARCH_HEENTRY, kvm_vcpu_arch, host_eentry);
+ OFFSET(KVM_ARCH_HPERCPU, kvm_vcpu_arch, host_percpu);
+
+ OFFSET(KVM_GPGD, kvm, arch.gpa_mm.pgd);
+ BLANK();
+}
diff --git a/arch/loongarch/kvm/switch.S b/arch/loongarch/kvm/switch.S
new file mode 100644
index 000000000000..7e55976c6fec
--- /dev/null
+++ b/arch/loongarch/kvm/switch.S
@@ -0,0 +1,299 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) 2020-2023 Loongson Technology Corporation Limited
+ */
+
+#include <linux/linkage.h>
+#include <asm/stackframe.h>
+#include <asm/asm.h>
+#include <asm/asmmacro.h>
+#include <asm/regdef.h>
+#include <asm/loongarch.h>
+#include <asm/export.h>
+
+#define PT_GPR_OFFSET(x) (PT_R0 + 8*x)
+ .text
The .text is redundant, you also marked it later.
Thanks, I will remove this .text
+
+.macro kvm_save_host_gpr base
+ .irp n,1,2,3,22,23,24,25,26,27,28,29,30,31
+ st.d $r\n, \base, PT_GPR_OFFSET(\n)
+ .endr
+.endm
+
+.macro kvm_restore_host_gpr base
+ .irp n,1,2,3,22,23,24,25,26,27,28,29,30,31
+ ld.d $r\n, \base, PT_GPR_OFFSET(\n)
+ .endr
+.endm
+
+/*
+ * prepare switch to guest, save host reg and restore guest reg.
+ * a2: kvm_vcpu_arch, don't touch it until 'ertn'
+ * t0, t1: temp register
+ */
+.macro kvm_switch_to_guest
+ /* set host excfg.VS=0, all exceptions share one exception entry */
+ csrrd t0, LOONGARCH_CSR_ECFG
+ bstrins.w t0, zero, CSR_ECFG_VS_SHIFT_END, CSR_ECFG_VS_SHIFT
+ csrwr t0, LOONGARCH_CSR_ECFG
+
+ /* Load up the new EENTRY */
+ ld.d t0, a2, KVM_ARCH_GEENTRY
+ csrwr t0, LOONGARCH_CSR_EENTRY
+
+ /* Set Guest ERA */
+ ld.d t0, a2, KVM_ARCH_GPC
+ csrwr t0, LOONGARCH_CSR_ERA
+
+ /* Save host PGDL */
+ csrrd t0, LOONGARCH_CSR_PGDL
+ st.d t0, a2, KVM_ARCH_HPGD
+
+ /* Switch to kvm */
+ ld.d t1, a2, KVM_VCPU_KVM - KVM_VCPU_ARCH
+
+ /* Load guest PGDL */
+ lu12i.w t0, KVM_GPGD
+ srli.w t0, t0, 12
I think the "lu12i.w+srli.w" can be replaced by "li.w t0, KVM_GPGD" .
Thanks, I will use "li.w t0, KVM_GPGD" to simplify the two instructions.
+ ldx.d t0, t1, t0
+ csrwr t0, LOONGARCH_CSR_PGDL
+
+ /* Mix GID and RID */
+ csrrd t1, LOONGARCH_CSR_GSTAT
+ bstrpick.w t1, t1, CSR_GSTAT_GID_SHIFT_END, CSR_GSTAT_GID_SHIFT
+ csrrd t0, LOONGARCH_CSR_GTLBC
+ bstrins.w t0, t1, CSR_GTLBC_TGID_SHIFT_END, CSR_GTLBC_TGID_SHIFT
+ csrwr t0, LOONGARCH_CSR_GTLBC
+
+ /*
+ * Switch to guest:
+ * GSTAT.PGM = 1, ERRCTL.ISERR = 0, TLBRPRMD.ISTLBR = 0
+ * ertn
+ */
+
+ /*
+ * Enable intr in root mode with future ertn so that host interrupt
+ * can be responsed during VM runs
+ * guest crmd comes from separate gcsr_CRMD register
+ */
+ ori t0, zero, CSR_PRMD_PIE
li.w t0, CSR_PRMD_PIE
Thanks for your advice, and I think it need not to replace it with
"li.w" there, as it has the same meaning with "ori" instruction, and
"ori" instruction is simple and clear enough. The same as the following
"move" instructions. What do you think of it.
+ csrxchg t0, t0, LOONGARCH_CSR_PRMD
+
+ /* Set PVM bit to setup ertn to guest context */
+ ori t0, zero, CSR_GSTAT_PVM
li.w t0, ...
+ csrxchg t0, t0, LOONGARCH_CSR_GSTAT
+
+ /* Load Guest gprs */
+ ld.d ra, a2, (KVM_ARCH_GGPR + 8 * REG_RA)
+ ld.d tp, a2, (KVM_ARCH_GGPR + 8 * REG_TP)
+ ld.d sp, a2, (KVM_ARCH_GGPR + 8 * REG_SP)
+ ld.d a0, a2, (KVM_ARCH_GGPR + 8 * REG_A0)
+ ld.d a1, a2, (KVM_ARCH_GGPR + 8 * REG_A1)
+ ld.d a3, a2, (KVM_ARCH_GGPR + 8 * REG_A3)
+ ld.d a4, a2, (KVM_ARCH_GGPR + 8 * REG_A4)
+ ld.d a5, a2, (KVM_ARCH_GGPR + 8 * REG_A5)
+ ld.d a6, a2, (KVM_ARCH_GGPR + 8 * REG_A6)
+ ld.d a7, a2, (KVM_ARCH_GGPR + 8 * REG_A7)
+ ld.d t0, a2, (KVM_ARCH_GGPR + 8 * REG_T0)
+ ld.d t1, a2, (KVM_ARCH_GGPR + 8 * REG_T1)
+ ld.d t2, a2, (KVM_ARCH_GGPR + 8 * REG_T2)
+ ld.d t3, a2, (KVM_ARCH_GGPR + 8 * REG_T3)
+ ld.d t4, a2, (KVM_ARCH_GGPR + 8 * REG_T4)
+ ld.d t5, a2, (KVM_ARCH_GGPR + 8 * REG_T5)
+ ld.d t6, a2, (KVM_ARCH_GGPR + 8 * REG_T6)
+ ld.d t7, a2, (KVM_ARCH_GGPR + 8 * REG_T7)
+ ld.d t8, a2, (KVM_ARCH_GGPR + 8 * REG_T8)
+ ld.d u0, a2, (KVM_ARCH_GGPR + 8 * REG_U0)
+ ld.d fp, a2, (KVM_ARCH_GGPR + 8 * REG_FP)
+ ld.d s0, a2, (KVM_ARCH_GGPR + 8 * REG_S0)
+ ld.d s1, a2, (KVM_ARCH_GGPR + 8 * REG_S1)
+ ld.d s2, a2, (KVM_ARCH_GGPR + 8 * REG_S2)
+ ld.d s3, a2, (KVM_ARCH_GGPR + 8 * REG_S3)
+ ld.d s4, a2, (KVM_ARCH_GGPR + 8 * REG_S4)
+ ld.d s5, a2, (KVM_ARCH_GGPR + 8 * REG_S5)
+ ld.d s6, a2, (KVM_ARCH_GGPR + 8 * REG_S6)
+ ld.d s7, a2, (KVM_ARCH_GGPR + 8 * REG_S7)
+ ld.d s8, a2, (KVM_ARCH_GGPR + 8 * REG_S8)
Why not use .irp which like kvm_save/restore_host_gpr do here?
Thanks, I will add kvm_restore_guest_gprs fucntion and use .irp
statement to replace these instructions.
+ /* Load KVM_ARCH register */
+ ld.d a2, a2, (KVM_ARCH_GGPR + 8 * REG_A2)
+
+ ertn
+.endm
+
+ /*
+ * exception entry for general exception from guest mode
+ * - IRQ is disabled
+ * - kernel privilege in root mode
+ * - page mode keep unchanged from previous prmd in root mode
+ * - Fixme: tlb exception cannot happen since registers
relative with TLB
+ * - is still in guest mode, such as pgd table/vmid
registers etc,
+ * - will fix with hw page walk enabled in future
+ * load kvm_vcpu from reserved CSR KVM_VCPU_KS, and save a2 to
KVM_TEMP_KS
+ */
+ .section .text
here.
+ .cfi_sections .debug_frame
+SYM_CODE_START(kvm_vector_entry)
+ csrwr a2, KVM_TEMP_KS
+ csrrd a2, KVM_VCPU_KS
+ addi.d a2, a2, KVM_VCPU_ARCH
+
+ /* After save gprs, free to use any gpr */
+ st.d ra, a2, (KVM_ARCH_GGPR + 8 * REG_RA)
+ st.d tp, a2, (KVM_ARCH_GGPR + 8 * REG_TP)
+ st.d sp, a2, (KVM_ARCH_GGPR + 8 * REG_SP)
+ st.d a0, a2, (KVM_ARCH_GGPR + 8 * REG_A0)
+ st.d a1, a2, (KVM_ARCH_GGPR + 8 * REG_A1)
+ st.d a3, a2, (KVM_ARCH_GGPR + 8 * REG_A3)
+ st.d a4, a2, (KVM_ARCH_GGPR + 8 * REG_A4)
+ st.d a5, a2, (KVM_ARCH_GGPR + 8 * REG_A5)
+ st.d a6, a2, (KVM_ARCH_GGPR + 8 * REG_A6)
+ st.d a7, a2, (KVM_ARCH_GGPR + 8 * REG_A7)
+ st.d t0, a2, (KVM_ARCH_GGPR + 8 * REG_T0)
+ st.d t1, a2, (KVM_ARCH_GGPR + 8 * REG_T1)
+ st.d t2, a2, (KVM_ARCH_GGPR + 8 * REG_T2)
+ st.d t3, a2, (KVM_ARCH_GGPR + 8 * REG_T3)
+ st.d t4, a2, (KVM_ARCH_GGPR + 8 * REG_T4)
+ st.d t5, a2, (KVM_ARCH_GGPR + 8 * REG_T5)
+ st.d t6, a2, (KVM_ARCH_GGPR + 8 * REG_T6)
+ st.d t7, a2, (KVM_ARCH_GGPR + 8 * REG_T7)
+ st.d t8, a2, (KVM_ARCH_GGPR + 8 * REG_T8)
+ st.d u0, a2, (KVM_ARCH_GGPR + 8 * REG_U0)
+ st.d fp, a2, (KVM_ARCH_GGPR + 8 * REG_FP)
+ st.d s0, a2, (KVM_ARCH_GGPR + 8 * REG_S0)
+ st.d s1, a2, (KVM_ARCH_GGPR + 8 * REG_S1)
+ st.d s2, a2, (KVM_ARCH_GGPR + 8 * REG_S2)
+ st.d s3, a2, (KVM_ARCH_GGPR + 8 * REG_S3)
+ st.d s4, a2, (KVM_ARCH_GGPR + 8 * REG_S4)
+ st.d s5, a2, (KVM_ARCH_GGPR + 8 * REG_S5)
+ st.d s6, a2, (KVM_ARCH_GGPR + 8 * REG_S6)
+ st.d s7, a2, (KVM_ARCH_GGPR + 8 * REG_S7)
+ st.d s8, a2, (KVM_ARCH_GGPR + 8 * REG_S8)
Why not .irp ... ?
Thanks, I will add kvm_save_guest_gprs fucntion and use .irp statement
to replace these instructions.
+ /* Save guest a2 */
+ csrrd t0, KVM_TEMP_KS
+ st.d t0, a2, (KVM_ARCH_GGPR + 8 * REG_A2)
+
+ /* a2: kvm_vcpu_arch, a1 is free to use */
+ csrrd s1, KVM_VCPU_KS
+ ld.d s0, s1, KVM_VCPU_RUN
+
+ csrrd t0, LOONGARCH_CSR_ESTAT
+ st.d t0, a2, KVM_ARCH_HESTAT
+ csrrd t0, LOONGARCH_CSR_ERA
+ st.d t0, a2, KVM_ARCH_GPC
+ csrrd t0, LOONGARCH_CSR_BADV
+ st.d t0, a2, KVM_ARCH_HBADV
+ csrrd t0, LOONGARCH_CSR_BADI
+ st.d t0, a2, KVM_ARCH_HBADI
+
+ /* Restore host excfg.VS */
+ csrrd t0, LOONGARCH_CSR_ECFG
+ ld.d t1, a2, KVM_ARCH_HECFG
+ or t0, t0, t1
+ csrwr t0, LOONGARCH_CSR_ECFG
+
+ /* Restore host eentry */
+ ld.d t0, a2, KVM_ARCH_HEENTRY
+ csrwr t0, LOONGARCH_CSR_EENTRY
+
+ /* restore host pgd table */
+ ld.d t0, a2, KVM_ARCH_HPGD
+ csrwr t0, LOONGARCH_CSR_PGDL
+
+ /*
+ * Disable PGM bit to enter root mode by default with next ertn
+ */
+ ori t0, zero, CSR_GSTAT_PVM
li.w
+ csrxchg zero, t0, LOONGARCH_CSR_GSTAT
+ /*
+ * Clear GTLBC.TGID field
+ * 0: for root tlb update in future tlb instr
+ * others: for guest tlb update like gpa to hpa in future tlb
instr
+ */
+ csrrd t0, LOONGARCH_CSR_GTLBC
+ bstrins.w t0, zero, CSR_GTLBC_TGID_SHIFT_END,
CSR_GTLBC_TGID_SHIFT
+ csrwr t0, LOONGARCH_CSR_GTLBC
+ ld.d tp, a2, KVM_ARCH_HTP
+ ld.d sp, a2, KVM_ARCH_HSP
+ /* restore per cpu register */
+ ld.d u0, a2, KVM_ARCH_HPERCPU
+ addi.d sp, sp, -PT_SIZE
+
+ /* Prepare handle exception */
+ or a0, s0, zero
move
+ or a1, s1, zero
move
+ ld.d t8, a2, KVM_ARCH_HANDLE_EXIT
+ jirl ra, t8, 0
+
+ or a2, s1, zero
move
+ addi.d a2, a2, KVM_VCPU_ARCH
+
+ /* resume host when ret <= 0 */
+ bge zero, a0, ret_to_host
blez a0, ret_to_host
I think we need not to change it, as they have the same meaning.
+
+ /*
+ * return to guest
+ * save per cpu register again, maybe switched to another cpu
+ */
+ st.d u0, a2, KVM_ARCH_HPERCPU
+
+ /* Save kvm_vcpu to kscratch */
+ csrwr s1, KVM_VCPU_KS
+ kvm_switch_to_guest
+
+ret_to_host:
+ ld.d a2, a2, KVM_ARCH_HSP
+ addi.d a2, a2, -PT_SIZE
+ kvm_restore_host_gpr a2
+ jr ra
+SYM_CODE_END(kvm_vector_entry)
+kvm_vector_entry_end:
+
+/*
+ * int kvm_enter_guest(struct kvm_run *run, struct kvm_vcpu *vcpu)
+ *
+ * @register_param:
+ * a0: kvm_run* run
+ * a1: kvm_vcpu* vcpu
+ */
+SYM_FUNC_START(kvm_enter_guest)
+ /* allocate space in stack bottom */
+ addi.d a2, sp, -PT_SIZE
+ /* save host gprs */
+ kvm_save_host_gpr a2
+
+ /* save host crmd,prmd csr to stack */
+ csrrd a3, LOONGARCH_CSR_CRMD
+ st.d a3, a2, PT_CRMD
+ csrrd a3, LOONGARCH_CSR_PRMD
+ st.d a3, a2, PT_PRMD
+
+ addi.d a2, a1, KVM_VCPU_ARCH
+ st.d sp, a2, KVM_ARCH_HSP
+ st.d tp, a2, KVM_ARCH_HTP
+ /* Save per cpu register */
+ st.d u0, a2, KVM_ARCH_HPERCPU
+
+ /* Save kvm_vcpu to kscratch */
+ csrwr a1, KVM_VCPU_KS
+ kvm_switch_to_guest
+SYM_FUNC_END(kvm_enter_guest)
+kvm_enter_guest_end:
+
+ .section ".rodata"
+SYM_DATA(kvm_vector_size, .quad kvm_vector_entry_end -
kvm_vector_entry)
+SYM_DATA(kvm_enter_guest_size, .quad kvm_enter_guest_end -
kvm_enter_guest)
Not restore .text section. Although I'd like to use .pushsection ... .pop
after marked kvm_..._end, you can put them after kvm_restore_fpu, too.
Jinyang
Thanks, I will put these code after kvm_restore_fpu.
Thanks
Tianrui Zhao
+
+
+SYM_FUNC_START(kvm_save_fpu)
+ fpu_save_csr a0 t1
+ fpu_save_double a0 t1
+ fpu_save_cc a0 t1 t2
+ jr ra
+SYM_FUNC_END(kvm_save_fpu)
+
+SYM_FUNC_START(kvm_restore_fpu)
+ fpu_restore_double a0 t1
+ fpu_restore_csr a0 t1
+ fpu_restore_cc a0 t1 t2
+ jr ra
+SYM_FUNC_END(kvm_restore_fpu)