lkml.org 
[lkml]   [2018]   [Aug]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v13 07/13] x86/sgx: Add data structures for tracking the EPC pages
From
Date
On 08/27/2018 11:53 AM, Jarkko Sakkinen wrote:
> Add data structures to track Enclave Page Cache (EPC) pages. EPC is
> divided into multiple banks (1-N) of which addresses and sizes can be
> enumerated with CPUID by the OS.
>
> On NUMA systems a node can have at most bank. A bank can be at most part of
> two nodes. SGX supports both nodes with a single memory controller and also
> sub-cluster nodes with severals memory controllers on a single die.
>
> Signed-off-by: Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>
> Co-developed-by: Serge Ayoun <serge.ayoun@intel.com>
> Co-developed-by: Sean Christopherson <sean.j.christopherson@intel.com>
> Signed-off-by: Serge Ayoun <serge.ayoun@intel.com>
> Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
> ---
> arch/x86/include/asm/sgx.h | 60 ++++++++++++++++++
> arch/x86/kernel/cpu/intel_sgx.c | 106 +++++++++++++++++++++++++++++++-
> 2 files changed, 164 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/include/asm/sgx.h b/arch/x86/include/asm/sgx.h
> index 2130e639ab49..17b7b3aa66bf 100644
> --- a/arch/x86/include/asm/sgx.h
> +++ b/arch/x86/include/asm/sgx.h
> @@ -4,9 +4,69 @@
> #ifndef _ASM_X86_SGX_H
> #define _ASM_X86_SGX_H
>
> +#include <linux/bitops.h>
> +#include <linux/err.h>
> +#include <linux/rwsem.h>
> #include <linux/types.h>
> +#include <asm/sgx_arch.h>
> +#include <asm/asm.h>
> +
> +#define SGX_MAX_EPC_BANKS 8

This is _still_ missing a meaningful description of what a bank is and
whether it is a hardware or software structure.

It would also help us to determine whether your bit packing below is
really required.

> +struct sgx_epc_page {
> + unsigned long desc;
> + struct list_head list;
> +};
> +
> +struct sgx_epc_bank {
> + unsigned long pa;
> + void *va;
> + unsigned long size;

Please add units. size could be bytes or pages, or who knows what. I
can't tell you how many bugs I've tripped over in the past from simple
unit conversions

> + struct sgx_epc_page *pages_data;
> + struct sgx_epc_page **pages;
> + unsigned long free_cnt;
> + spinlock_t lock;
> +};
>
> extern bool sgx_enabled;
> extern bool sgx_lc_enabled;
> +extern struct sgx_epc_bank sgx_epc_banks[SGX_MAX_EPC_BANKS];
> +
> +/*
> + * enum sgx_epc_page_desc - defines bits and masks for an EPC page's desc

Why are you bothering packing these bits? This seems a rather
convoluted way to store two integers.

> +static __init int sgx_init_epc_bank(u64 addr, u64 size, unsigned long index,
> + struct sgx_epc_bank *bank)
> +{
> + unsigned long nr_pages = size >> PAGE_SHIFT;
> + struct sgx_epc_page *pages_data;
> + unsigned long i;
> + void *va;
> +
> + va = ioremap_cache(addr, size);
> + if (!va)
> + return -ENOMEM;
> +
> + pages_data = kcalloc(nr_pages, sizeof(struct sgx_epc_page), GFP_KERNEL);
> + if (!pages_data)
> + goto out_iomap;

This looks like you're roughly limited by the page allocator to a bank
size of ~1.4GB which seems kinda small. Is this really OK?

> + bank->pages = kcalloc(nr_pages, sizeof(struct sgx_epc_page *),
> + GFP_KERNEL);
> + if (!bank->pages)
> + goto out_pdata;
> +
> + for (i = 0; i < nr_pages; i++) {
> + bank->pages[i] = &pages_data[i];
> + bank->pages[i]->desc = (addr + (i << PAGE_SHIFT)) | index;
> + }
> +
> + bank->pa = addr;
> + bank->size = size;
> + bank->va = va;
> + bank->free_cnt = nr_pages;
> + bank->pages_data = pages_data;
> + spin_lock_init(&bank->lock);
> + return 0;
> +out_pdata:
> + kfree(pages_data);
> +out_iomap:
> + iounmap(va);
> + return -ENOMEM;
> +}
> +
> +static __init void sgx_page_cache_teardown(void)
> +{
> + struct sgx_epc_bank *bank;
> + int i;
> +
> + for (i = 0; i < sgx_nr_epc_banks; i++) {
> + bank = &sgx_epc_banks[i];
> + iounmap((void *)bank->va);
> + kfree(bank->pages);
> + kfree(bank->pages_data);
> + }
> +}
> +
> +static inline u64 sgx_combine_bank_regs(u64 low, u64 high)
> +{
> + return (low & 0xFFFFF000) + ((high & 0xFFFFF) << 32);
> +}

-ENOCOMMENT for a rather weird looking calculation

> +static __init int sgx_page_cache_init(void)
> +{
> + u32 eax, ebx, ecx, edx;
> + u64 pa, size;
> + int ret;
> + int i;
> +
> + for (i = 0; i < SGX_MAX_EPC_BANKS; i++) {
> + cpuid_count(SGX_CPUID, 2 + i, &eax, &ebx, &ecx, &edx);
> + if (!(eax & 0xF))
> + break;

So, we have random data coming out of a random CPUID leaf being called
'eax' and then being tested against a random hard-coded mask. This
seems rather unfortunate for someone trying to understand the code. Can
we do better?

> + pa = sgx_combine_bank_regs(eax, ebx);
> + size = sgx_combine_bank_regs(ecx, edx);
> + pr_info("EPC bank 0x%llx-0x%llx\n", pa, pa + size - 1);
> + ret = sgx_init_epc_bank(pa, size, i, &sgx_epc_banks[i]);
> + if (ret) {
> + sgx_page_cache_teardown();
> + return ret;
> + }

So if one bank fails, we tear down all banks, yet leave sgx_nr_epc_banks
incremented? That sounds troublesome.

> + sgx_nr_epc_banks++;
> + }
> +
> + if (!sgx_nr_epc_banks) {
> + pr_err("There are zero EPC banks.\n");
> + return -ENODEV;
> + }
> +
> + return 0;
> +}

Does this support hot-addition of a bank? If not, why not?

\
 
 \ /
  Last update: 2018-08-27 23:08    [W:0.390 / U:0.192 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site