diff mbox

[intel-sgx-kernel-dev,v6,8/8] intel_sgx: add LRU algorithm to page swapping

Message ID 20161204184044.21031-9-jarkko.sakkinen@linux.intel.com (mailing list archive)
State New, archived
Headers show

Commit Message

Jarkko Sakkinen Dec. 4, 2016, 6:40 p.m. UTC 
From: Sean Christopherson <sean.j.christopherson@intel.com>

Test and clear the A bit of an EPC page when isolating pages during EPC
page swap.  Move accessed pages to the end of the load list instead of
the eviction list, i.e. isolate only those pages that have not been
accessed since the last time the swapping flow was run.

This basic LRU algorithm yields a significant improvement in throughput
when the system is under a heavy EPC pressure.

This patch is based on the code originally written Serge Ayoun for the
out-of-tree driver with a bug fix that the EPC pages are moved always to
the end of the list if they have been lately accessed. For unknown
reason a version of out-of-tree driver slipped into the Github, which
this regression.

Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Reviewed-by: Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>
Tested-by: Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>
---
 drivers/platform/x86/intel_sgx.h            |  2 +-
 drivers/platform/x86/intel_sgx_ioctl.c      |  1 +
 drivers/platform/x86/intel_sgx_page_cache.c | 59 +++++++++++++++++++++--------
 drivers/platform/x86/intel_sgx_vma.c        |  1 +
 4 files changed, 46 insertions(+), 17 deletions(-)
diff mbox

Patch

diff --git a/drivers/platform/x86/intel_sgx.h b/drivers/platform/x86/intel_sgx.h
index add3565..b659b71 100644
--- a/drivers/platform/x86/intel_sgx.h
+++ b/drivers/platform/x86/intel_sgx.h
@@ -193,7 +193,7 @@  long sgx_compat_ioctl(struct file *filep, unsigned int cmd, unsigned long arg);
 #endif
 
 /* Utility functions */
-
+int sgx_test_and_clear_young(struct sgx_encl_page *page, struct sgx_encl *encl);
 void *sgx_get_epc_page(struct sgx_epc_page *entry);
 void sgx_put_epc_page(void *epc_page_vaddr);
 struct page *sgx_get_backing(struct sgx_encl *encl,
diff --git a/drivers/platform/x86/intel_sgx_ioctl.c b/drivers/platform/x86/intel_sgx_ioctl.c
index ab0a4a3..53fa510 100644
--- a/drivers/platform/x86/intel_sgx_ioctl.c
+++ b/drivers/platform/x86/intel_sgx_ioctl.c
@@ -296,6 +296,7 @@  static bool sgx_process_add_page_req(struct sgx_add_page_req *req)
 	}
 
 	encl_page->epc_page = epc_page;
+	sgx_test_and_clear_young(encl_page, encl);
 	list_add_tail(&encl_page->load_list, &encl->load_list);
 
 	mutex_unlock(&encl->lock);
diff --git a/drivers/platform/x86/intel_sgx_page_cache.c b/drivers/platform/x86/intel_sgx_page_cache.c
index f2a2ed1..8a411b3 100644
--- a/drivers/platform/x86/intel_sgx_page_cache.c
+++ b/drivers/platform/x86/intel_sgx_page_cache.c
@@ -77,6 +77,42 @@  static unsigned int sgx_nr_high_pages;
 struct task_struct *ksgxswapd_tsk;
 static DECLARE_WAIT_QUEUE_HEAD(ksgxswapd_waitq);
 
+
+static int sgx_test_and_clear_young_cb(pte_t *ptep, pgtable_t token,
+				       unsigned long addr, void *data)
+{
+	pte_t pte;
+	int ret;
+
+	ret = pte_young(*ptep);
+	if (ret) {
+		pte = pte_mkold(*ptep);
+		set_pte_at((struct mm_struct *)data, addr, ptep, pte);
+	}
+
+	return ret;
+}
+
+/**
+ * sgx_test_and_clear_young() - Test and reset the accessed bit
+ * @page:	enclave EPC page to be tested for recent access
+ * @encl:	enclave which owns @page
+ *
+ * Checks the Access (A) bit from the PTE corresponding to the
+ * enclave page and clears it.  Returns 1 if the page has been
+ * recently accessed and 0 if not.
+ */
+int sgx_test_and_clear_young(struct sgx_encl_page *page, struct sgx_encl *encl)
+{
+	struct vm_area_struct *vma = sgx_find_vma(encl, page->addr);
+
+	if (!vma)
+		return 0;
+
+	return apply_to_page_range(vma->vm_mm, page->addr, PAGE_SIZE,
+				   sgx_test_and_clear_young_cb, vma->vm_mm);
+}
+
 static struct sgx_tgid_ctx *sgx_isolate_tgid_ctx(unsigned long nr_to_scan)
 {
 	struct sgx_tgid_ctx *ctx = NULL;
@@ -166,7 +202,8 @@  static void sgx_isolate_pages(struct sgx_encl *encl,
 					 struct sgx_encl_page,
 					 load_list);
 
-		if (!(entry->flags & SGX_ENCL_PAGE_RESERVED)) {
+		if (!sgx_test_and_clear_young(entry, encl) &&
+		    !(entry->flags & SGX_ENCL_PAGE_RESERVED)) {
 			entry->flags |= SGX_ENCL_PAGE_RESERVED;
 			list_move_tail(&entry->load_list, dst);
 		} else {
@@ -267,19 +304,6 @@  static void sgx_write_pages(struct sgx_encl *encl, struct list_head *src)
 
 	entry = list_first_entry(src, struct sgx_encl_page, load_list);
 
-	if (!sgx_pin_mm(encl)) {
-		while (!list_empty(src)) {
-			entry = list_first_entry(src, struct sgx_encl_page,
-						 load_list);
-			list_del(&entry->load_list);
-			mutex_lock(&encl->lock);
-			sgx_free_encl_page(entry, encl, 0);
-			mutex_unlock(&encl->lock);
-		}
-
-		return;
-	}
-
 	mutex_lock(&encl->lock);
 
 	/* EBLOCK */
@@ -345,8 +369,6 @@  static void sgx_write_pages(struct sgx_encl *encl, struct list_head *src)
 	}
 
 	mutex_unlock(&encl->lock);
-
-	sgx_unpin_mm(encl);
 }
 
 static void sgx_swap_pages(unsigned long nr_to_scan)
@@ -363,9 +385,14 @@  static void sgx_swap_pages(unsigned long nr_to_scan)
 	if (!encl)
 		goto out;
 
+	if (!sgx_pin_mm(encl))
+		goto out_enclave;
+
 	sgx_isolate_pages(encl, &cluster, nr_to_scan);
 	sgx_write_pages(encl, &cluster);
+	sgx_unpin_mm(encl);
 
+out_enclave:
 	kref_put(&encl->refcount, sgx_encl_release);
 out:
 	kref_put(&ctx->refcount, sgx_tgid_ctx_release);
diff --git a/drivers/platform/x86/intel_sgx_vma.c b/drivers/platform/x86/intel_sgx_vma.c
index 4515cc3c..1cf5ba9 100644
--- a/drivers/platform/x86/intel_sgx_vma.c
+++ b/drivers/platform/x86/intel_sgx_vma.c
@@ -262,6 +262,7 @@  static struct sgx_encl_page *sgx_vma_do_fault(struct vm_area_struct *vma,
 	/* Do not free */
 	epc_page = NULL;
 
+	sgx_test_and_clear_young(entry, encl);
 	list_add_tail(&entry->load_list, &encl->load_list);
 out:
 	mutex_unlock(&encl->lock);