On Thu, Nov 02, 2017 at 01:10:29PM -0700, Sean Christopherson wrote:
> On Tue, 2017-10-10 at 17:32 +0300, Jarkko Sakkinen wrote:
> > +static int sgx_dev_init(struct device *parent, bool locked_msrs)
> > +{
> > + struct sgx_context *sgx_dev;
> > + unsigned int eax, ebx, ecx, edx;
> > + unsigned long pa;
> > + unsigned long size;
> > + int ret;
> > + int i;
> > +
> > + pr_info("intel_sgx: " DRV_DESCRIPTION " v" DRV_VERSION "\n");
> > +
> > + sgx_dev = sgxm_ctx_alloc(parent);
> > +
> > + sgx_locked_msrs = locked_msrs;
> > +
> > + cpuid_count(SGX_CPUID, SGX_CPUID_CAPABILITIES, &eax, &ebx, &ecx,
> > &edx);
> > + /* Only allow misc bits supported by the driver. */
> > + sgx_misc_reserved = ~ebx | SGX_MISC_RESERVED_MASK;
> > +#ifdef CONFIG_X86_64
> > + sgx_encl_size_max_64 = 1ULL << ((edx >> 8) & 0xFF);
> > +#endif
> > + sgx_encl_size_max_32 = 1ULL << (edx & 0xFF);
> > +
> > + if (boot_cpu_has(X86_FEATURE_OSXSAVE)) {
> > + cpuid_count(SGX_CPUID, SGX_CPUID_ATTRIBUTES, &eax, &ebx,
> > &ecx,
> > + &edx);
> > + sgx_xfrm_mask = (((u64)edx) << 32) + (u64)ecx;
> > +
> > + for (i = 2; i < 64; i++) {
> > + cpuid_count(0x0D, i, &eax, &ebx, &ecx, &edx);
> > + if ((1 << i) & sgx_xfrm_mask)
> > + sgx_xsave_size_tbl[i] = eax + ebx;
> > + }
> > + }
> > +
> > + for (i = 0; i < SGX_MAX_EPC_BANKS; i++) {
> > + cpuid_count(SGX_CPUID, i + SGX_CPUID_EPC_BANKS, &eax, &ebx,
> > + &ecx, &edx);
> > + if (!(eax & 0xf))
> > + break;
> > +
> > + pa = ((u64)(ebx & 0xfffff) << 32) + (u64)(eax & 0xfffff000);
> > + size = ((u64)(edx & 0xfffff) << 32) + (u64)(ecx &
> > 0xfffff000);
> > +
> > + dev_info(parent, "EPC bank 0x%lx-0x%lx\n", pa, pa + size);
> > +
> > + sgx_epc_banks[i].pa = pa;
> > + sgx_epc_banks[i].size = size;
> > + }
> > +
> > + sgx_nr_epc_banks = i;
> > +
> > + for (i = 0; i < sgx_nr_epc_banks; i++) {
> > +#ifdef CONFIG_X86_64
> > + sgx_epc_banks[i].va = (unsigned long)
> > + ioremap_cache(sgx_epc_banks[i].pa,
> > + sgx_epc_banks[i].size);
> > + if (!sgx_epc_banks[i].va) {
> > + sgx_nr_epc_banks = i;
> > + ret = -ENOMEM;
> > + goto out_iounmap;
> > + }
> > +#endif
> > + ret = sgx_add_epc_bank(sgx_epc_banks[i].pa,
> > + sgx_epc_banks[i].size, i);
> > + if (ret) {
> > + sgx_nr_epc_banks = i + 1;
> > + goto out_iounmap;
> > + }
> > + }
> > +
> > + ret = sgx_page_cache_init();
> > + if (ret)
> > + goto out_iounmap;
> > +
> > + sgx_add_page_wq = alloc_workqueue("intel_sgx-add-page-wq",
> > + WQ_UNBOUND | WQ_FREEZABLE, 1);
> > + if (!sgx_add_page_wq) {
> > + pr_err("intel_sgx: alloc_workqueue() failed\n");
> > + ret = -ENOMEM;
> > + goto out_iounmap;
> > + }
> > +
> > + ret = cdev_device_add(&sgx_dev->cdev, &sgx_dev->dev);
> > + if (ret)
> > + goto out_workqueue;
> > +
> > + return 0;
> > +out_workqueue:
> > + destroy_workqueue(sgx_add_page_wq);
> > +out_iounmap:
>
> sgx_page_cache_teardown() should be called here, else ksgxswapd and the list of
> EPC pages will leak.
>
Thanks.
> > +#ifdef CONFIG_X86_64
> > + for (i = 0; i < sgx_nr_epc_banks; i++)
> > + iounmap((void *)sgx_epc_banks[i].va);
> > +#endif
> > + return ret;
> > +}
>
> ...
>
> > +int sgx_add_epc_bank(resource_size_t start, unsigned long size, int bank)
> > +{
> > + unsigned long i;
> > + struct sgx_epc_page *new_epc_page, *entry;
> > + struct list_head *parser, *temp;
> > +
> > + for (i = 0; i < size; i += PAGE_SIZE) {
> > + new_epc_page = kzalloc(sizeof(*new_epc_page), GFP_KERNEL);
> > + if (!new_epc_page)
> > + goto err_freelist;
> > + new_epc_page->pa = (start + i) | bank;
> > +
> > + spin_lock(&sgx_free_list_lock);
> > + list_add_tail(&new_epc_page->list, &sgx_free_list);
> > + sgx_nr_total_epc_pages++;
> > + sgx_nr_free_pages++;
> > + spin_unlock(&sgx_free_list_lock);
> > + }
> > +
> > + return 0;
> > +err_freelist:
> > + list_for_each_safe(parser, temp, &sgx_free_list) {
> > + spin_lock(&sgx_free_list_lock);
> > + entry = list_entry(parser, struct sgx_epc_page, list);
> > + list_del(&entry->list);
> > + spin_unlock(&sgx_free_list_lock);
> > + kfree(entry);
> > + }
> > + return -ENOMEM;
> > +}
>
> Freeing the entire list on failure does not seem like the appropriate behavior
> for this helper func, e.g. the list should be purged by sgx_page_cache_teardown.
> Buffering the new pages into a local list and only splicing said list into the
> global list upon success is more inline with what I would expect from a helper
> func, and also only requires a single lock/unlock.
>
> diff --git drivers/platform/x86/intel_sgx/sgx_page_cache.c
> drivers/platform/x86/intel_sgx/sgx_page_cache.c
> index f8883d24692a..38496e6296f1 100644
> --- drivers/platform/x86/intel_sgx/sgx_page_cache.c
> +++ drivers/platform/x86/intel_sgx/sgx_page_cache.c
> @@ -397,7 +397,8 @@ int sgx_add_epc_bank(resource_size_t start, unsigned long
> size, int bank)
> {
> unsigned long i;
> struct sgx_epc_page *new_epc_page, *entry;
> - struct list_head *parser, *temp;
> + LIST_HEAD(epc_pages);
>
> for (i = 0; i < size; i += PAGE_SIZE) {
> new_epc_page = kzalloc(sizeof(*new_epc_page), GFP_KERNEL);
> @@ -405,22 +406,19 @@ int sgx_add_epc_bank(resource_size_t start, unsigned long
> size, int bank)
> goto err_freelist;
> new_epc_page->pa = (start + i) | bank;
>
> - spin_lock(&sgx_free_list_lock);
> - list_add_tail(&new_epc_page->list, &sgx_free_list);
> - sgx_nr_total_epc_pages++;
> - sgx_nr_free_pages++;
> - spin_unlock(&sgx_free_list_lock);
> + list_add_tail(&new_epc_page->list, &epc_pages);
> + nr_pages++;
> }
>
> + spin_lock(&sgx_free_list_lock);
> + list_splice_tail(&epc_pages, &sgx_free_list);
> + sgx_nr_total_epc_pages += nr_pages;
> + sgx_nr_free_pages += nr_pages;
> + spin_unlock(&sgx_free_list_lock);
> return 0;
> err_freelist:
> - list_for_each_safe(parser, temp, &sgx_free_list) {
> - spin_lock(&sgx_free_list_lock);
> - entry = list_entry(parser, struct sgx_epc_page, list);
> - list_del(&entry->list);
> - spin_unlock(&sgx_free_list_lock);
> + list_for_each_entry(entry, &sgx_free_list, list)
> kfree(entry);
> - }
> return -ENOMEM;
> }
This makes sense to me and I think it makes also sense to move bunch of
stuff from sgx_dev_init() to sgx_page_cache_init().
/Jarkko
