)() On Thu, Sep 9, 2021 at 9:26 AM Brijesh Singh <brijesh.singh@xxxxxxx> wrote: > > > > On 9/9/21 9:54 AM, Peter Gonda wrote: > > On Fri, Aug 20, 2021 at 9:22 AM Brijesh Singh <brijesh.singh@xxxxxxx> wrote: > >> > >> The SNP guest request message header contains a message count. The > >> message count is used while building the IV. The PSP firmware increments > >> the message count by 1, and expects that next message will be using the > >> incremented count. The snp_msg_seqno() helper will be used by driver to > >> get the message sequence counter used in the request message header, > >> and it will be automatically incremented after the request is successful. > >> The incremented value is saved in the secrets page so that the kexec'ed > >> kernel knows from where to begin. > >> > >> Signed-off-by: Brijesh Singh <brijesh.singh@xxxxxxx> > >> --- > >> arch/x86/kernel/sev.c | 79 +++++++++++++++++++++++++++++++++++++++ > >> include/linux/sev-guest.h | 37 ++++++++++++++++++ > >> 2 files changed, 116 insertions(+) > >> > >> diff --git a/arch/x86/kernel/sev.c b/arch/x86/kernel/sev.c > >> index 319a40fc57ce..f42cd5a8e7bb 100644 > >> --- a/arch/x86/kernel/sev.c > >> +++ b/arch/x86/kernel/sev.c > >> @@ -51,6 +51,8 @@ static struct ghcb boot_ghcb_page __bss_decrypted __aligned(PAGE_SIZE); > >> */ > >> static struct ghcb __initdata *boot_ghcb; > >> > >> +static u64 snp_secrets_phys; > >> + > >> /* #VC handler runtime per-CPU data */ > >> struct sev_es_runtime_data { > >> struct ghcb ghcb_page; > >> @@ -2030,6 +2032,80 @@ bool __init handle_vc_boot_ghcb(struct pt_regs *regs) > >> halt(); > >> } > >> > >> +static struct snp_secrets_page_layout *snp_map_secrets_page(void) > >> +{ > >> + u16 __iomem *secrets; > >> + > >> + if (!snp_secrets_phys || !sev_feature_enabled(SEV_SNP)) > >> + return NULL; > >> + > >> + secrets = ioremap_encrypted(snp_secrets_phys, PAGE_SIZE); > >> + if (!secrets) > >> + return NULL; > >> + > >> + return (struct snp_secrets_page_layout *)secrets; > >> +} > >> + > >> +static inline u64 snp_read_msg_seqno(void) > >> +{ > >> + struct snp_secrets_page_layout *layout; > >> + u64 count; > >> + > >> + layout = snp_map_secrets_page(); > >> + if (!layout) > >> + return 0; > >> + > >> + /* Read the current message sequence counter from secrets pages */ > >> + count = readl(&layout->os_area.msg_seqno_0); > >> + > >> + iounmap(layout); > >> + > >> + /* The sequence counter must begin with 1 */ > >> + if (!count) > >> + return 1; > >> + > >> + return count + 1; > >> +} > >> + > >> +u64 snp_msg_seqno(void) > >> +{ > >> + u64 count = snp_read_msg_seqno(); > >> + > >> + if (unlikely(!count)) > >> + return 0; > >> + > >> + /* > >> + * The message sequence counter for the SNP guest request is a > >> + * 64-bit value but the version 2 of GHCB specification defines a > >> + * 32-bit storage for the it. > >> + */ > >> + if (count >= UINT_MAX) > >> + return 0; > >> + > >> + return count; > >> +} > >> +EXPORT_SYMBOL_GPL(snp_msg_seqno); > > > > Do we need some sort of get sequence number, then ack that sequence > > number was used API? Taking your host changes in Part2 V5 as an > > example. If 'snp_setup_guest_buf' fails the given sequence number is > > never actually used by a message to the PSP. So the guest will have > > the wrong current sequence number, an off by 1 error, right? > > > > The sequence number should be incremented only after the command is > successful. In this particular case the next caller should not get the > updated sequence number. > > Having said so, there is a bug in current code that will cause us to > increment the sequence number on failure. I notice it last week and have > it fixed in v6 wip branch. > > int snp_issue_guest_request(....) > { > > ..... > ..... > > ret = sev_es_ghcb_hv_call(ghcb, NULL, id, input->req_gpa, input->resp_gpa); > if (ret) > goto e_put; > > if (ghcb->save.sw_exit_info_2) { > ... > ... > > ret = -EIO; > goto e_put; /** THIS WAS MISSING */ > } > > /* The command was successful, increment the sequence counter. */ > snp_gen_msg_seqno(); > e_put: > .... > } > > Does this address your concern? So the 'snp_msg_seqno()' call in 'enc_payload' will not increment the counter, its only incremented on 'snp_gen_msg_seqno()'? If thats correct, that addresses my first concern. > > > > Also it seems like there is a concurrency error waiting to happen > > here. If 2 callers call snp_msg_seqno() before either actually places > > a call to the PSP, if the first caller's request doesn't reach the PSP > > before the second caller's request both calls will fail. And again I > > think the sequence numbers in the guest will be incorrect and > > unrecoverable. > > > > So far, the only user for the snp_msg_seqno() is the attestation driver. > And the driver is designed to serialize the vmgexit request and thus we > should not run into concurrence issue. That seems a little dangerous as any module new code or out-of-tree module could use this function thus revealing this race condition right? Could we at least have a comment on these functions (snp_msg_seqno and snp_gen_msg_seqno) noting this? > > >> + > >> +static void snp_gen_msg_seqno(void) > >> +{ > >> + struct snp_secrets_page_layout *layout; > >> + u64 count; > >> + > >> + layout = snp_map_secrets_page(); > >> + if (!layout) > >> + return; > >> + > >> + /* > >> + * The counter is also incremented by the PSP, so increment it by 2 > >> + * and save in secrets page. > >> + */ > >> + count = readl(&layout->os_area.msg_seqno_0); > >> + count += 2; > >> + > >> + writel(count, &layout->os_area.msg_seqno_0); > >> + iounmap(layout); > >> +} > >> + > >> int snp_issue_guest_request(int type, struct snp_guest_request_data *input, unsigned long *fw_err) > >> { > >> struct ghcb_state state; > >> @@ -2077,6 +2153,9 @@ int snp_issue_guest_request(int type, struct snp_guest_request_data *input, unsi > >> ret = -EIO; > >> } > >> > >> + /* The command was successful, increment the sequence counter */ > >> + snp_gen_msg_seqno(); > >> + > >> e_put: > >> __sev_put_ghcb(&state); > >> e_restore_irq: > >> diff --git a/include/linux/sev-guest.h b/include/linux/sev-guest.h > >> index 24dd17507789..16b6af24fda7 100644 > >> --- a/include/linux/sev-guest.h > >> +++ b/include/linux/sev-guest.h > >> @@ -20,6 +20,41 @@ enum vmgexit_type { > >> GUEST_REQUEST_MAX > >> }; > >> > >> +/* > >> + * The secrets page contains 96-bytes of reserved field that can be used by > >> + * the guest OS. The guest OS uses the area to save the message sequence > >> + * number for each VMPCK. > >> + * > >> + * See the GHCB spec section Secret page layout for the format for this area. > >> + */ > >> +struct secrets_os_area { > >> + u32 msg_seqno_0; > >> + u32 msg_seqno_1; > >> + u32 msg_seqno_2; > >> + u32 msg_seqno_3; > >> + u64 ap_jump_table_pa; > >> + u8 rsvd[40]; > >> + u8 guest_usage[32]; > >> +} __packed; > >> + > >> +#define VMPCK_KEY_LEN 32 > >> + > >> +/* See the SNP spec for secrets page format */ > >> +struct snp_secrets_page_layout { > >> + u32 version; > >> + u32 imien : 1, > >> + rsvd1 : 31; > >> + u32 fms; > >> + u32 rsvd2; > >> + u8 gosvw[16]; > >> + u8 vmpck0[VMPCK_KEY_LEN]; > >> + u8 vmpck1[VMPCK_KEY_LEN]; > >> + u8 vmpck2[VMPCK_KEY_LEN]; > >> + u8 vmpck3[VMPCK_KEY_LEN]; > >> + struct secrets_os_area os_area; > >> + u8 rsvd3[3840]; > >> +} __packed; > >> + > >> /* > >> * The error code when the data_npages is too small. The error code > >> * is defined in the GHCB specification. > >> @@ -36,6 +71,7 @@ struct snp_guest_request_data { > >> #ifdef CONFIG_AMD_MEM_ENCRYPT > >> int snp_issue_guest_request(int vmgexit_type, struct snp_guest_request_data *input, > >> unsigned long *fw_err); > >> +u64 snp_msg_seqno(void); > >> #else > >> > >> static inline int snp_issue_guest_request(int type, struct snp_guest_request_data *input, > >> @@ -43,6 +79,7 @@ static inline int snp_issue_guest_request(int type, struct snp_guest_request_dat > >> { > >> return -ENODEV; > >> } > >> +static inline u64 snp_msg_seqno(void) { return 0; } > >> > >> #endif /* CONFIG_AMD_MEM_ENCRYPT */ > >> #endif /* __LINUX_SEV_GUEST_H__ */ > >> -- > >> 2.17.1 > >> > >>