[lustre-devel] [PATCH 22/29] lustre: osc: fall back to vmalloc for large RPCs

[James Simmons]

From: Andreas Dilger <adilger at whamcloud.com>

For large RPC sizes (16MB+) the page array (4096 brw_page) can
become very large (128KB+ with fscrypt) and should fall back to
vmalloc() if kmalloc() fails due to memory fragmentation.

The mdc/mdt allocations are currently limited to 1MB for readdir
RPCs, but it doesn't hurt to prepare them for larger RPCs from
clients in the future if this limit is increased.

Fixes: 5965de814e91 ("staging: lustre: rpc: increase bulk size")
WC-bug-id: https://jira.whamcloud.com/browse/LU-13212
Lustre-commit: 037a9e2cf6d5b8d6f ("LU-13212 osc: fall back to vmalloc for large RPCs")
Signed-off-by: Andreas Dilger <adilger at whamcloud.com>
Reviewed-on: https://review.whamcloud.com/43281
Reviewed-by: Bobi Jam <bobijam at hotmail.com>
Reviewed-by: Wang Shilong <wshilong at whamcloud.com>
Reviewed-by: James Simmons <jsimmons at infradead.org>
Reviewed-by: Oleg Drokin <green at whamcloud.com>
Signed-off-by: James Simmons <jsimmons at infradead.org>
---
 fs/lustre/mdc/mdc_request.c     | 5 +++--
 fs/lustre/mgc/mgc_request.c     | 5 +++--
 fs/lustre/obdecho/echo_client.c | 5 +++--
 fs/lustre/osc/osc_request.c     | 5 +++--
 4 files changed, 12 insertions(+), 8 deletions(-)

diff --git a/fs/lustre/mdc/mdc_request.c b/fs/lustre/mdc/mdc_request.c
index 6ac3a39..fb9971f 100644
--- a/fs/lustre/mdc/mdc_request.c
+++ b/fs/lustre/mdc/mdc_request.c
@@ -1312,7 +1312,8 @@ static int mdc_read_page_remote(void *data, struct page *page0)
 	fid = &op_data->op_fid1;
 	LASSERT(inode);
 
-	page_pool = kcalloc(max_pages, sizeof(page), GFP_NOFS);
+	page_pool = kvmalloc_array(max_pages, sizeof(page),
+				   GFP_KERNEL | __GFP_ZERO);
 	if (page_pool) {
 		page_pool[0] = page0;
 	} else {
@@ -1381,7 +1382,7 @@ static int mdc_read_page_remote(void *data, struct page *page0)
 	}
 
 	if (page_pool != &page0)
-		kfree(page_pool);
+		kvfree(page_pool);
 
 	return rc;
 }
diff --git a/fs/lustre/mgc/mgc_request.c b/fs/lustre/mgc/mgc_request.c
index f115479..4b60056a 100644
--- a/fs/lustre/mgc/mgc_request.c
+++ b/fs/lustre/mgc/mgc_request.c
@@ -1350,7 +1350,8 @@ static int mgc_process_recover_log(struct obd_device *obd,
 	if (cfg->cfg_last_idx == 0) /* the first time */
 		nrpages = CONFIG_READ_NRPAGES_INIT;
 
-	pages = kcalloc(nrpages, sizeof(*pages), GFP_KERNEL);
+	pages = kvmalloc_array(nrpages, sizeof(*pages),
+			       GFP_KERNEL | __GFP_ZERO);
 	if (!pages) {
 		rc = -ENOMEM;
 		goto out;
@@ -1474,7 +1475,7 @@ static int mgc_process_recover_log(struct obd_device *obd,
 				break;
 			__free_page(pages[i]);
 		}
-		kfree(pages);
+		kvfree(pages);
 	}
 	return rc;
 }
diff --git a/fs/lustre/obdecho/echo_client.c b/fs/lustre/obdecho/echo_client.c
index 3bee0c2..270226b 100644
--- a/fs/lustre/obdecho/echo_client.c
+++ b/fs/lustre/obdecho/echo_client.c
@@ -1302,7 +1302,8 @@ static int echo_client_kbrw(struct echo_device *ed, int rw, struct obdo *oa,
 	if (!pga)
 		return -ENOMEM;
 
-	pages = kcalloc(npages, sizeof(*pages), GFP_NOFS);
+	pages = kvmalloc_array(npages, sizeof(*pages),
+			       GFP_KERNEL | __GFP_ZERO);
 	if (!pages) {
 		kfree(pga);
 		return -ENOMEM;
@@ -1355,7 +1356,7 @@ static int echo_client_kbrw(struct echo_device *ed, int rw, struct obdo *oa,
 		__free_page(pgp->pg);
 	}
 	kfree(pga);
-	kfree(pages);
+	kvfree(pages);
 	return rc;
 }
 
diff --git a/fs/lustre/osc/osc_request.c b/fs/lustre/osc/osc_request.c
index 1bd0d2c..973c504 100644
--- a/fs/lustre/osc/osc_request.c
+++ b/fs/lustre/osc/osc_request.c
@@ -2323,7 +2323,7 @@ static void sort_brw_pages(struct brw_page **array, int num)
 static void osc_release_ppga(struct brw_page **ppga, u32 count)
 {
 	LASSERT(ppga);
-	kfree(ppga);
+	kvfree(ppga);
 }
 
 static int brw_interpret(const struct lu_env *env,
@@ -2523,7 +2523,8 @@ int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
 	if (mem_tight)
 		mpflag = memalloc_noreclaim_save();
 
-	pga = kcalloc(page_count, sizeof(*pga), GFP_NOFS);
+	pga = kvmalloc_array(page_count, sizeof(*pga),
+			     GFP_KERNEL | __GFP_ZERO);
 	if (!pga) {
 		rc = -ENOMEM;
 		goto out;
-- 
1.8.3.1