Re: [tip: x86/urgent] x86/sev: Do not touch VMSA pages during SNP guest memory kdump

* tip-bot2 for Ashish Kalra tip-bot2@linutronix.de wrote:

The following commit has been merged into the x86/urgent branch of tip:

Commit-ID: d2062cc1b1c367d5d019f595ef860159e1301351 Gitweb: https://git.kernel.org/tip/d2062cc1b1c367d5d019f595ef860159e1301351 Author: Ashish Kalra ashish.kalra@amd.com AuthorDate: Mon, 28 Apr 2025 21:41:51 Committer: Borislav Petkov (AMD) bp@alien8.de CommitterDate: Tue, 13 May 2025 19:40:44 +02:00

x86/sev: Do not touch VMSA pages during SNP guest memory kdump

When kdump is running makedumpfile to generate vmcore and dump SNP guest memory it touches the VMSA page of the vCPU executing kdump.

It then results in unrecoverable #NPF/RMP faults as the VMSA page is marked busy/in-use when the vCPU is running and subsequently a causes guest softlockup/hang.

s/subsequently a causes /subsequently causes

Additionally, other APs may be halted in guest mode and their VMSA pages are marked busy and touching these VMSA pages during guest memory dump will also cause #NPF.

Issue AP_DESTROY GHCB calls on other APs to ensure they are kicked out of guest mode and then clear the VMSA bit on their VMSA pages.

If the vCPU running kdump is an AP, mark it's VMSA page as offline to ensure that makedumpfile excludes that page while dumping guest memory.

s/mark it's VMSA page /mark its VMSA page

Fixes: 3074152e56c9 ("x86/sev: Convert shared memory back to private on kexec") Signed-off-by: Ashish Kalra ashish.kalra@amd.com Signed-off-by: Borislav Petkov (AMD) bp@alien8.de Reviewed-by: Pankaj Gupta pankaj.gupta@amd.com Reviewed-by: Tom Lendacky thomas.lendacky@amd.com Tested-by: Srikanth Aithal sraithal@amd.com Cc: stable@vger.kernel.org Link: https://lore.kernel.org/20250428214151.155464-1-Ashish.Kalra@amd.com

---

arch/x86/coco/sev/core.c | 244 ++++++++++++++++++++++++-------------- 1 file changed, 158 insertions(+), 86 deletions(-)

diff --git a/arch/x86/coco/sev/core.c b/arch/x86/coco/sev/core.c index b0c1a7a..41060ba 100644 --- a/arch/x86/coco/sev/core.c +++ b/arch/x86/coco/sev/core.c @@ -959,6 +959,102 @@ void snp_accept_memory(phys_addr_t start, phys_addr_t end) set_pages_state(vaddr, npages, SNP_PAGE_STATE_PRIVATE); } +static int vmgexit_ap_control(u64 event, struct sev_es_save_area *vmsa, u32 apic_id) +{

• bool create = event != SVM_VMGEXIT_AP_DESTROY;
• struct ghcb_state state;
• unsigned long flags;
• struct ghcb *ghcb;
• int ret = 0;
• local_irq_save(flags);
• ghcb = __sev_get_ghcb(&state);
• vc_ghcb_invalidate(ghcb);
• if (create)

• ghcb_set_rax(ghcb, vmsa->sev_features);
  

• ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_AP_CREATION);
• ghcb_set_sw_exit_info_1(ghcb,

• 		((u64)apic_id << 32)	|
  

• 		((u64)snp_vmpl << 16)	|
  

• 		event);
  

• ghcb_set_sw_exit_info_2(ghcb, __pa(vmsa));
• sev_es_wr_ghcb_msr(__pa(ghcb));
• VMGEXIT();
• if (!ghcb_sw_exit_info_1_is_valid(ghcb) ||

•    lower_32_bits(ghcb->save.sw_exit_info_1)) {
  

• pr_err("SNP AP %s error\n", (create ? "CREATE" : "DESTROY"));
  

• ret = -EINVAL;
  

• }
• __sev_put_ghcb(&state);
• local_irq_restore(flags);
• return ret;

+}

+static int snp_set_vmsa(void *va, void *caa, int apic_id, bool make_vmsa) +{

• int ret;
• if (snp_vmpl) {

• struct svsm_call call = {};
  

• unsigned long flags;
  

• local_irq_save(flags);
  

• call.caa = this_cpu_read(svsm_caa);
  

• call.rcx = __pa(va);
  

• if (make_vmsa) {
  

• 	/* Protocol 0, Call ID 2 */
  

• 	call.rax = SVSM_CORE_CALL(SVSM_CORE_CREATE_VCPU);
  

• 	call.rdx = __pa(caa);
  

This can probably use svsm_caa_pa instead of __pa(), like sev_es_init_vc_handling() does, see below.

• 	call.r8  = apic_id;
  

• } else {
  

• 	/* Protocol 0, Call ID 3 */
  

• 	call.rax = SVSM_CORE_CALL(SVSM_CORE_DELETE_VCPU);
  

• }
  

• ret = svsm_perform_call_protocol(&call);
  

• local_irq_restore(flags);
  

• } else {

• /*
  

•  * If the kernel runs at VMPL0, it can change the VMSA
  

•  * bit for a page using the RMPADJUST instruction.
  

•  * However, for the instruction to succeed it must
  

•  * target the permissions of a lesser privileged (higher
  

•  * numbered) VMPL level, so use VMPL1.
  

•  */
  

• u64 attrs = 1;
  

• if (make_vmsa)
  

• 	attrs |= RMPADJUST_VMSA_PAGE_BIT;
  

• ret = rmpadjust((unsigned long)va, RMP_PG_SIZE_4K, attrs);
  

• }
• return ret;

+}

+static void snp_cleanup_vmsa(struct sev_es_save_area *vmsa, int apic_id) +{

• int err;
• err = snp_set_vmsa(vmsa, NULL, apic_id, false);
• if (err)

• pr_err("clear VMSA page failed (%u), leaking page\n", err);
  

• else

• free_page((unsigned long)vmsa);
  

So the argument types here are really messy:

- We pass in a 'struct sev_es_save_area *vmsa' to snp_cleanup_vmsa(), which passes it down to snp_set_vmsa() as a void *, where it's force-type-cast to 'unsigned long' ...

- While within snp_cleanup_vmsa() we also force-cast it to 'unsigned long' yet again.

It would be much cleaner to do a single, obvious force-cast to a virtual address type within snp_cleanup_vmsa():

unsigned long vmsa_va = (unsigned long)vmsa;

And change snp_set_vmsa()'s parameter to 'unsigned long vmsa_va', to get rid of a lot of forced/dangerous type conversions.

Plus the handling of 'caa' pointers it really messy AFAICS:

- alloc_runtime_data() calculates svsm_caa_pa physical addresses for each CPU:

per_cpu(svsm_caa_pa, cpu) = __pa(caa);

Which is used by sev_es_init_vc_handling():

call.rcx = this_cpu_read(svsm_caa_pa);

But snp_set_vmsa() calculates the physical address *again* instead of using svsm_caa_pa:

call.caa = this_cpu_read(svsm_caa); ... call.rdx = __pa(caa);

Same for snp_set_vmsa():

call.caa = this_cpu_read(svsm_caa); call.rcx = __pa(va);

Why? Either this is something subtle and undocumented, or at minimum this unnecessarily complicates the code and creates inconsistent patterns of implementing the same functionality.

+}

static void set_pte_enc(pte_t *kpte, int level, void *va) { struct pte_enc_desc d = { @@ -1055,6 +1151,65 @@ void snp_kexec_begin(void) pr_warn("Failed to stop shared<->private conversions\n"); } +/*

• • Shutdown all APs except the one handling kexec/kdump and clearing
• • the VMSA tag on AP's VMSA pages as they are not being used as
• • VMSA page anymore.

s/Shutdown Shut down

'shutdown' is a noun, the verb is 'to shut down'.

• */

+static void shutdown_all_aps(void) +{

• struct sev_es_save_area *vmsa;
• int apic_id, this_cpu, cpu;
• this_cpu = get_cpu();
• /*

• * APs are already in HLT loop when enc_kexec_finish() callback
  

• * is invoked.
  

• */
  

• for_each_present_cpu(cpu) {

• vmsa = per_cpu(sev_vmsa, cpu);
  

• /*
  

•  * The BSP or offlined APs do not have guest allocated VMSA
  

•  * and there is no need  to clear the VMSA tag for this page.
  

Whitespace noise:

s/ / /

•  */
  

• if (!vmsa)
  

• 	continue;
  

• /*
  

•  * Cannot clear the VMSA tag for the currently running vCPU.
  

•  */
  

• if (this_cpu == cpu) {
  

• 	unsigned long pa;
  

• 	struct page *p;
  

• 	pa = __pa(vmsa);
  

• 	/*
  

• 	 * Mark the VMSA page of the running vCPU as offline
  

• 	 * so that is excluded and not touched by makedumpfile
  

• 	 * while generating vmcore during kdump.
  

s/so that is excluded /so that it is excluded

• 	 */
  

• 	p = pfn_to_online_page(pa >> PAGE_SHIFT);
  

• 	if (p)
  

• 		__SetPageOffline(p);
  

• 	continue;
  

• }
  

• apic_id = cpuid_to_apicid[cpu];
  

• /*
  

•  * Issue AP destroy to ensure AP gets kicked out of guest mode
  

•  * to allow using RMPADJUST to remove the VMSA tag on it's
  

•  * VMSA page.
  

s/on it's VMSA page /on its VMSA page

•  */
  

• vmgexit_ap_control(SVM_VMGEXIT_AP_DESTROY, vmsa, apic_id);
  

• snp_cleanup_vmsa(vmsa, apic_id);
  

Boris, please don't rush these SEV patches without proper review first! ;-)

Thanks,

Ingo