[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <20230820202715.29006-6-decui@microsoft.com>
Date: Sun, 20 Aug 2023 13:27:11 -0700
From: Dexuan Cui <decui@...rosoft.com>
To: ak@...ux.intel.com, arnd@...db.de, bp@...en8.de,
brijesh.singh@....com, dan.j.williams@...el.com,
dave.hansen@...el.com, dave.hansen@...ux.intel.com,
haiyangz@...rosoft.com, hpa@...or.com, jane.chu@...cle.com,
kirill.shutemov@...ux.intel.com, kys@...rosoft.com,
linux-hyperv@...r.kernel.org, luto@...nel.org, mingo@...hat.com,
peterz@...radead.org, rostedt@...dmis.org,
sathyanarayanan.kuppuswamy@...ux.intel.com, seanjc@...gle.com,
tglx@...utronix.de, tony.luck@...el.com, wei.liu@...nel.org,
Jason@...c4.com, nik.borisov@...e.com, mikelley@...rosoft.com
Cc: x86@...nel.org, linux-kernel@...r.kernel.org,
linux-arch@...r.kernel.org, Tianyu.Lan@...rosoft.com,
rick.p.edgecombe@...el.com, andavis@...hat.com, mheslin@...hat.com,
vkuznets@...hat.com, xiaoyao.li@...el.com,
Dexuan Cui <decui@...rosoft.com>
Subject: [PATCH v2 5/9] Drivers: hv: vmbus: Support >64 VPs for a fully enlightened TDX/SNP VM
Don't set *this_cpu_ptr(hyperv_pcpu_input_arg) before the function
set_memory_decrypted() returns, otherwise we run into this ticky issue:
For a fully enlightened TDX/SNP VM, in hv_common_cpu_init(),
*this_cpu_ptr(hyperv_pcpu_input_arg) is an encrypted page before
the set_memory_decrypted() returns.
When such a VM has more than 64 VPs, if the hyperv_pcpu_input_arg is not
NULL, hv_common_cpu_init() -> set_memory_decrypted() -> ... ->
cpa_flush() -> on_each_cpu() -> ... -> hv_send_ipi_mask() -> ... ->
__send_ipi_mask_ex() tries to call hv_do_fast_hypercall16() with the
hyperv_pcpu_input_arg as the hypercall input page, which must be a
decrypted page in such a VM, but the page is still encrypted at this
point, and a fatal fault is triggered.
Fix the issue by setting *this_cpu_ptr(hyperv_pcpu_input_arg) after
set_memory_decrypted(): if the hyperv_pcpu_input_arg is NULL,
__send_ipi_mask_ex() returns HV_STATUS_INVALID_PARAMETER immediately,
and hv_send_ipi_mask() falls back to orig_apic.send_IPI_mask(),
which can use x2apic_send_IPI_all(), which may be slightly slower than
the hypercall but still works correctly in such a VM.
Signed-off-by: Dexuan Cui <decui@...rosoft.com>
---
Changes in v2: None
drivers/hv/hv_common.c | 30 +++++++++++++++++++++++-------
1 file changed, 23 insertions(+), 7 deletions(-)
diff --git a/drivers/hv/hv_common.c b/drivers/hv/hv_common.c
index 897bbb96f4118..4c858e1636da7 100644
--- a/drivers/hv/hv_common.c
+++ b/drivers/hv/hv_common.c
@@ -360,6 +360,7 @@ int hv_common_cpu_init(unsigned int cpu)
u64 msr_vp_index;
gfp_t flags;
int pgcount = hv_root_partition ? 2 : 1;
+ void *mem;
int ret;
/* hv_cpu_init() can be called with IRQs disabled from hv_resume() */
@@ -372,25 +373,40 @@ int hv_common_cpu_init(unsigned int cpu)
* allocated if this CPU was previously online and then taken offline
*/
if (!*inputarg) {
- *inputarg = kmalloc(pgcount * HV_HYP_PAGE_SIZE, flags);
- if (!(*inputarg))
+ mem = kmalloc(pgcount * HV_HYP_PAGE_SIZE, flags);
+ if (!mem)
return -ENOMEM;
if (hv_root_partition) {
outputarg = (void **)this_cpu_ptr(hyperv_pcpu_output_arg);
- *outputarg = (char *)(*inputarg) + HV_HYP_PAGE_SIZE;
+ *outputarg = (char *)mem + HV_HYP_PAGE_SIZE;
}
if (hv_isolation_type_en_snp() || hv_isolation_type_tdx()) {
- ret = set_memory_decrypted((unsigned long)*inputarg, pgcount);
+ ret = set_memory_decrypted((unsigned long)mem, pgcount);
if (ret) {
- /* It may be unsafe to free *inputarg */
- *inputarg = NULL;
+ /* It may be unsafe to free 'mem' */
return ret;
}
- memset(*inputarg, 0x00, pgcount * PAGE_SIZE);
+ memset(mem, 0x00, pgcount * HV_HYP_PAGE_SIZE);
}
+
+ /*
+ * In a fully enlightened TDX/SNP VM with more than 64 VPs, if
+ * hyperv_pcpu_input_arg is not NULL, set_memory_decrypted() ->
+ * ... -> cpa_flush()-> ... -> __send_ipi_mask_ex() tries to
+ * use hyperv_pcpu_input_arg as the hypercall input page, which
+ * must be a decrypted page in such a VM, but the page is still
+ * encrypted before set_memory_decrypted() returns. Fix this by
+ * setting *inputarg after the above set_memory_decrypted(): if
+ * hyperv_pcpu_input_arg is NULL, __send_ipi_mask_ex() returns
+ * HV_STATUS_INVALID_PARAMETER immediately, and the function
+ * hv_send_ipi_mask() falls back to orig_apic.send_IPI_mask(),
+ * which may be slightly slower than the hypercall, but still
+ * works correctly in such a VM.
+ */
+ *inputarg = mem;
}
msr_vp_index = hv_get_register(HV_REGISTER_VP_INDEX);
--
2.25.1
Powered by blists - more mailing lists