Re: [PATCH 2/2] support kdump when AMD secure memory encryption is active

From: lijiang
Date: Wed May 16 2018 - 20:15:16 EST


å 2018å05æ16æ 04:18, Tom Lendacky åé:
> On 5/14/2018 8:51 PM, Lianbo Jiang wrote:
>> When sme enabled on AMD server, we also need to support kdump. Because
>> the memory is encrypted in the first kernel, we will remap the old memory
>> encrypted to the second kernel(crash kernel), and sme is also enabled in
>> the second kernel, otherwise the old memory encrypted can not be decrypted.
>> Because simply changing the value of a C-bit on a page will not
>> automatically encrypt the existing contents of a page, and any data in the
>> page prior to the C-bit modification will become unintelligible. A page of
>> memory that is marked encrypted will be automatically decrypted when read
>> from DRAM and will be automatically encrypted when written to DRAM.
>>
>> For the kdump, it is necessary to distinguish whether the memory is
>> encrypted. Furthermore, we should also know which part of the memory is
>> encrypted or decrypted. We will appropriately remap the memory according
>> to the specific situation in order to tell cpu how to deal with the data(
>> encrypted or unencrypted). For example, when sme enabled, if the old memory
>> is encrypted, we will remap the old memory in encrypted way, which will
>> automatically decrypt the old memory encrypted when we read those data from
>> the remapping address.
>>
>> ----------------------------------------------
>> | first-kernel | second-kernel | kdump support |
>> | (mem_encrypt=on|off) | (yes|no) |
>> |--------------+---------------+---------------|
>> | on | on | yes |
>> | off | off | yes |
>> | on | off | no |
>> | off | on | no |
>> |______________|_______________|_______________|
>>
>> Signed-off-by: Lianbo Jiang <lijiang@xxxxxxxxxx>
>> ---
>> arch/x86/include/asm/dmi.h | 14 +++++++++++++-
>> arch/x86/kernel/acpi/boot.c | 8 ++++++++
>> arch/x86/kernel/crash_dump_64.c | 27 +++++++++++++++++++++++++++
>> drivers/acpi/tables.c | 14 +++++++++++++-
>> drivers/iommu/amd_iommu_init.c | 9 ++++++++-
>> fs/proc/vmcore.c | 36 +++++++++++++++++++++++++++++++-----
>> include/linux/crash_dump.h | 4 ++++
>> kernel/kexec_core.c | 12 ++++++++++++
>> 8 files changed, 116 insertions(+), 8 deletions(-)
>>
>> diff --git a/arch/x86/include/asm/dmi.h b/arch/x86/include/asm/dmi.h
>> index 0ab2ab2..a5663b4 100644
>> --- a/arch/x86/include/asm/dmi.h
>> +++ b/arch/x86/include/asm/dmi.h
>> @@ -7,6 +7,10 @@
>>
>> #include <asm/io.h>
>> #include <asm/setup.h>
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>
> I don't think you need all of the #ifdef stuff throughout this
> patch. Everything should work just fine without it.
>
>> +#include <linux/crash_dump.h>
>> +#include <linux/mem_encrypt.h>
>> +#endif
>>
>> static __always_inline __init void *dmi_alloc(unsigned len)
>> {
>> @@ -14,7 +18,15 @@ static __always_inline __init void *dmi_alloc(unsigned len)
>> }
>>
>> /* Use early IO mappings for DMI because it's initialized early */
>> -#define dmi_early_remap early_memremap
>> +static __always_inline __init void *dmi_early_remap(resource_size_t
>> + phys_addr, unsigned long size)
>> +{
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>
> Again, no need for the #ifdef here. You should probably audit the
> code for all of these and truly determine if they are really needed.
>
>> + if (sme_active() && is_kdump_kernel())
>
> Use of sme_active() here is good since under SEV, this area will be
> encrypted.
>
>> + return early_memremap_decrypted(phys_addr, size);
>> +#endif
>> + return early_memremap(phys_addr, size);
>
> Instead of doing this, maybe it makes more sense to put this logic
> somewhere in the early_memremap() path. Possibly smarten up the
> early_memremap_pgprot_adjust() function with some kdump kernel
> related logic. Not sure it's possible, but would be nice since you
> have this logic in a couple of places.
>
>> +}
>> #define dmi_early_unmap early_memunmap
>> #define dmi_remap(_x, _l) memremap(_x, _l, MEMREMAP_WB)
>> #define dmi_unmap(_x) memunmap(_x)
>> diff --git a/arch/x86/kernel/acpi/boot.c b/arch/x86/kernel/acpi/boot.c
>> index 3b20607..354ad66 100644
>> --- a/arch/x86/kernel/acpi/boot.c
>> +++ b/arch/x86/kernel/acpi/boot.c
>> @@ -48,6 +48,10 @@
>> #include <asm/mpspec.h>
>> #include <asm/smp.h>
>> #include <asm/i8259.h>
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>> +#include <linux/crash_dump.h>
>> +#include <linux/mem_encrypt.h>
>> +#endif
>>
>> #include "sleep.h" /* To include x86_acpi_suspend_lowlevel */
>> static int __initdata acpi_force = 0;
>> @@ -124,6 +128,10 @@ void __init __iomem *__acpi_map_table(unsigned long phys, unsigned long size)
>> if (!phys || !size)
>> return NULL;
>>
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>> + if (sme_active() && is_kdump_kernel())
>> + return early_memremap_decrypted(phys, size);
>> +#endif
>
> Same as previous comment(s).
>
>> return early_memremap(phys, size);
>> }
>>
>> diff --git a/arch/x86/kernel/crash_dump_64.c b/arch/x86/kernel/crash_dump_64.c
>> index 4f2e077..2ef67fc 100644
>> --- a/arch/x86/kernel/crash_dump_64.c
>> +++ b/arch/x86/kernel/crash_dump_64.c
>> @@ -48,3 +48,30 @@ ssize_t copy_oldmem_page(unsigned long pfn, char *buf,
>> iounmap(vaddr);
>> return csize;
>> }
>> +
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>> +ssize_t copy_oldmem_page_encrypted(unsigned long pfn, char *buf,
>> + size_t csize, unsigned long offset, int userbuf)
>> +{
>> + void *vaddr;
>> +
>> + if (!csize)
>> + return 0;
>> +
>> + vaddr = ioremap_encrypted(pfn << PAGE_SHIFT, PAGE_SIZE);
>> + if (!vaddr)
>> + return -ENOMEM;
>> +
>> + if (userbuf) {
>> + if (copy_to_user(buf, vaddr + offset, csize)) {
>> + iounmap(vaddr);
>> + return -EFAULT;
>> + }
>> + } else
>> + memcpy(buf, vaddr + offset, csize);
>> +
>> + set_iounmap_nonlazy();
>> + iounmap(vaddr);
>> + return csize;
>> +}
>> +#endif
>
> This seems exactly the same as copy_oldmem_page() with the difference
> being the type of ioremap done. Might be better to make the code
> after the ioremap's a common piece of code that each of the copy_oldmem
> functions would call.
>
>> diff --git a/drivers/acpi/tables.c b/drivers/acpi/tables.c
>> index 849c4fb..6da9b0c 100644
>> --- a/drivers/acpi/tables.c
>> +++ b/drivers/acpi/tables.c
>> @@ -36,6 +36,10 @@
>> #include <linux/memblock.h>
>> #include <linux/initrd.h>
>> #include "internal.h"
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>> +#include <linux/crash_dump.h>
>> +#include <linux/mem_encrypt.h>
>> +#endif
>>
>> #ifdef CONFIG_ACPI_CUSTOM_DSDT
>> #include CONFIG_ACPI_CUSTOM_DSDT_FILE
>> @@ -566,7 +570,15 @@ void __init acpi_table_upgrade(void)
>> clen = size;
>> if (clen > MAP_CHUNK_SIZE - slop)
>> clen = MAP_CHUNK_SIZE - slop;
>> - dest_p = early_memremap(dest_addr & PAGE_MASK,
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>> + if (sme_active() && is_kdump_kernel())
>> + dest_p = early_memremap_decrypted(
>> + dest_addr & PAGE_MASK,
>> + clen + slop);
>> + else
>> +#endif
>> + dest_p = early_memremap(
>> + dest_addr & PAGE_MASK,
>
> So if the dest_addr (based off of acpi_tables_addr) was added to the e820
> map as an ACPI area (which it will be), then it would be mapped properly
> (in both SME and SEV) without needing the if/then/else.
>
>> clen + slop);
>> memcpy(dest_p + slop, src_p, clen);
>> early_memunmap(dest_p, clen + slop);
>> diff --git a/drivers/iommu/amd_iommu_init.c b/drivers/iommu/amd_iommu_init.c
>> index 904c575..8ecbddb 100644
>> --- a/drivers/iommu/amd_iommu_init.c
>> +++ b/drivers/iommu/amd_iommu_init.c
>> @@ -889,11 +889,18 @@ static bool copy_device_table(void)
>> }
>>
>> old_devtb_phys = entry & PAGE_MASK;
>> + if (sme_active() && is_kdump_kernel())
>
> Use mem_encrypt_active() here to cover both SME and SEV.
>
>> + old_devtb_phys = __sme_clr(old_devtb_phys);
>> if (old_devtb_phys >= 0x100000000ULL) {
>> pr_err("The address of old device table is above 4G, not trustworthy!\n");
>> return false;
>> }
>> - old_devtb = memremap(old_devtb_phys, dev_table_size, MEMREMAP_WB);
>> + if (sme_active() && is_kdump_kernel())
>> + old_devtb = ioremap_encrypted(old_devtb_phys,
>> + dev_table_size);
>> + else
>> + old_devtb = memremap(old_devtb_phys,
>> + dev_table_size, MEMREMAP_WB);
>
> What happens to the memremap here, does it fall back to the ioremap and
> end up getting mapped decrypted? It would be nice to do the right thing
> under the covers of memremap. Not sure what that would take, but it would
> keep the code nice and clean.
>
Thank you, Tom.
For this issue, it will copy the old content of device table entries, that is to
say, it will copy the device table from the first kernel to the second kernel in
kdump mode and the content of device table is encrypted, the address includes the
sme_me_mask. For the old memory encrypted, it is suitable to use ioremap_encrypted,
the kdump kernel can not directly access the first kernel's memory.

Lianbo
>> if (!old_devtb)
>> return false;
>>
>> diff --git a/fs/proc/vmcore.c b/fs/proc/vmcore.c
>> index a45f0af..316e2b0 100644
>> --- a/fs/proc/vmcore.c
>> +++ b/fs/proc/vmcore.c
>> @@ -25,6 +25,10 @@
>> #include <linux/uaccess.h>
>> #include <asm/io.h>
>> #include "internal.h"
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>> +#include <linux/mem_encrypt.h>
>> +#include <asm/pgtable.h>
>> +#endif
>>
>> /* List representing chunks of contiguous memory areas and their offsets in
>> * vmcore file.
>> @@ -86,7 +90,8 @@ static int pfn_is_ram(unsigned long pfn)
>>
>> /* Reads a page from the oldmem device from given offset. */
>> static ssize_t read_from_oldmem(char *buf, size_t count,
>> - u64 *ppos, int userbuf)
>> + u64 *ppos, int userbuf,
>> + bool encrypted)
>> {
>> unsigned long pfn, offset;
>> size_t nr_bytes;
>> @@ -108,8 +113,15 @@ static ssize_t read_from_oldmem(char *buf, size_t count,
>> if (pfn_is_ram(pfn) == 0)
>> memset(buf, 0, nr_bytes);
>> else {
>> - tmp = copy_oldmem_page(pfn, buf, nr_bytes,
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>> + if (encrypted)
>> + tmp = copy_oldmem_page_encrypted(pfn, buf,
>> + nr_bytes, offset, userbuf);
>> + else
>> +#endif
>> + tmp = copy_oldmem_page(pfn, buf, nr_bytes,
>> offset, userbuf);> +
>> if (tmp < 0)
>> return tmp;
>> }
>> @@ -143,7 +155,7 @@ void __weak elfcorehdr_free(unsigned long long addr)
>> */
>> ssize_t __weak elfcorehdr_read(char *buf, size_t count, u64 *ppos)
>> {
>> - return read_from_oldmem(buf, count, ppos, 0);
>> + return read_from_oldmem(buf, count, ppos, 0, false);
>
> For SEV, this will likely be encrypted, so you can probably replace the
> "false" with sev_active() so that under SME it is un-encrypted but under
> SEV it is encrypted. Where is the elfcorehdr stored? I wonder if it
> could be created as encrypted under SME and then you could actually remove
> the encrypted parameter from read_from_oldmem() and always map encrypted.
> If SME or SEV are active it will be mapped encrypted and if they aren't
> then it is mapped normally.
>
>> }
>>
>> /*
>> @@ -151,7 +163,11 @@ ssize_t __weak elfcorehdr_read(char *buf, size_t count, u64 *ppos)
>> */
>> ssize_t __weak elfcorehdr_read_notes(char *buf, size_t count, u64 *ppos)
>> {
>> - return read_from_oldmem(buf, count, ppos, 0);
>> + bool flag = false;
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>> + flag = sme_active();
>> +#endif
>> + return read_from_oldmem(buf, count, ppos, 0, flag);
>> }
>>
>> /*
>> @@ -161,6 +177,10 @@ int __weak remap_oldmem_pfn_range(struct vm_area_struct *vma,
>> unsigned long from, unsigned long pfn,
>> unsigned long size, pgprot_t prot)
>> {
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>> + if (sme_active())
>
> No need for the sme_active() check here, the encryption will be applied,
> for both SME and SEV, if memory encryption is active, otherwise it won't.
>
>> + prot = __pgprot(pgprot_val(prot) | _PAGE_ENC);
>
> prot = pgprot_encrypted(prot);
>
>> +#endif> return remap_pfn_range(vma, from, pfn, size, prot);
>> }
>>
>> @@ -188,6 +208,11 @@ static ssize_t __read_vmcore(char *buffer, size_t buflen, loff_t *fpos,
>> size_t tsz;
>> u64 start;
>> struct vmcore *m = NULL;
>> + bool sme_flag = false;
>> +
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>> + sme_flag = sme_active();> +#endif
>
> Probably just want mem_encrypt_active() here to get both SME and SEV
> cases mapped as encrypted.
>
> Thanks,
> Tom
>
>>
>> if (buflen == 0 || *fpos >= vmcore_size)
>> return 0;
>> @@ -235,7 +260,8 @@ static ssize_t __read_vmcore(char *buffer, size_t buflen, loff_t *fpos,
>> m->offset + m->size - *fpos,
>> buflen);
>> start = m->paddr + *fpos - m->offset;
>> - tmp = read_from_oldmem(buffer, tsz, &start, userbuf);
>> + tmp = read_from_oldmem(buffer, tsz, &start,
>> + userbuf, sme_flag);
>> if (tmp < 0)
>> return tmp;
>> buflen -= tsz;
>> diff --git a/include/linux/crash_dump.h b/include/linux/crash_dump.h
>> index f7ac2aa..024ae9e 100644
>> --- a/include/linux/crash_dump.h
>> +++ b/include/linux/crash_dump.h
>> @@ -25,6 +25,10 @@ extern int remap_oldmem_pfn_range(struct vm_area_struct *vma,
>>
>> extern ssize_t copy_oldmem_page(unsigned long, char *, size_t,
>> unsigned long, int);
>> +#ifdef CONFIG_AMD_MEM_ENCRYPT
>> +extern ssize_t copy_oldmem_page_encrypted(unsigned long, char *, size_t,
>> + unsigned long, int);
>> +#endif
>> void vmcore_cleanup(void);
>>
>> /* Architecture code defines this if there are other possible ELF
>> diff --git a/kernel/kexec_core.c b/kernel/kexec_core.c
>> index 20fef1a..3c22a9b 100644
>> --- a/kernel/kexec_core.c
>> +++ b/kernel/kexec_core.c
>> @@ -471,6 +471,16 @@ static struct page *kimage_alloc_crash_control_pages(struct kimage *image,
>> }
>> }
>>
>> + if (pages) {
>> + unsigned int count, i;
>> +
>> + pages->mapping = NULL;
>> + set_page_private(pages, order);
>> + count = 1 << order;
>> + for (i = 0; i < count; i++)
>> + SetPageReserved(pages + i);
>> + arch_kexec_post_alloc_pages(page_address(pages), 1 << order, 0);
>> + }
>> return pages;
>> }
>>
>> @@ -865,6 +875,7 @@ static int kimage_load_crash_segment(struct kimage *image,
>> result = -ENOMEM;
>> goto out;
>> }
>> + arch_kexec_post_alloc_pages(page_address(page), 1, 0);
>> ptr = kmap(page);
>> ptr += maddr & ~PAGE_MASK;
>> mchunk = min_t(size_t, mbytes,
>> @@ -882,6 +893,7 @@ static int kimage_load_crash_segment(struct kimage *image,
>> result = copy_from_user(ptr, buf, uchunk);
>> kexec_flush_icache_page(page);
>> kunmap(page);
>> + arch_kexec_pre_free_pages(page_address(page), 1);
>> if (result) {
>> result = -EFAULT;
>> goto out;
>>