Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Peilin Ye <yepeilin.cs@gmail.com>
To:     Alexander Viro <viro@zeniv.linux.org.uk>
Cc:     Peilin Ye <peilin.ye@bytedance.com>,
        Cong Wang <cong.wang@bytedance.com>,
        linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org,
        Peilin Ye <yepeilin.cs@gmail.com>
Subject: [PATCH v1] coredump: Use vmsplice_to_pipe() for pipes in dump_emit_page()
Date:   Fri, 28 Oct 2022 17:51:47 -0700
Message-Id: <20221029005147.2553-1-yepeilin.cs@gmail.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

From: Peilin Ye <peilin.ye@bytedance.com>

Currently, there is a copy for each page when dumping VMAs to pipe
handlers using dump_emit_page().  For example:

  fs/binfmt_elf.c:elf_core_dump()
      fs/coredump.c:dump_user_range()
                     :dump_emit_page()
        fs/read_write.c:__kernel_write_iter()
                fs/pipe.c:pipe_write()
             lib/iov_iter.c:copy_page_from_iter()

Use vmsplice_to_pipe() instead of __kernel_write_iter() to avoid this
copy for pipe handlers.

Tested by dumping a 40-GByte core into a simple handler that splice()s
from stdin to disk in a loop, PIPE_DEF_BUFFERS (16) pages at a time.

                              Before           After   Improved by
  Time to Completion   52.04 seconds   46.30 seconds        11.03%
  CPU Usage                   89.43%          84.90%         5.07%

Suggested-by: Cong Wang <cong.wang@bytedance.com>
Signed-off-by: Peilin Ye <peilin.ye@bytedance.com>
---
 fs/coredump.c          | 7 ++++++-
 fs/splice.c            | 4 ++--
 include/linux/splice.h | 2 ++
 3 files changed, 10 insertions(+), 3 deletions(-)

diff --git a/fs/coredump.c b/fs/coredump.c
index 7bad7785e8e6..a6ef406dee43 100644
--- a/fs/coredump.c
+++ b/fs/coredump.c
@@ -42,6 +42,7 @@
 #include <linux/timekeeping.h>
 #include <linux/sysctl.h>
 #include <linux/elf.h>
+#include <linux/splice.h>
 
 #include <linux/uaccess.h>
 #include <asm/mmu_context.h>
@@ -854,7 +855,11 @@ static int dump_emit_page(struct coredump_params *cprm, struct page *page)
 		return 0;
 	pos = file->f_pos;
 	iov_iter_bvec(&iter, WRITE, &bvec, 1, PAGE_SIZE);
-	n = __kernel_write_iter(cprm->file, &iter, &pos);
+
+	n = vmsplice_to_pipe(file, &iter, 0);
+	if (n == -EBADF)
+		n = __kernel_write_iter(cprm->file, &iter, &pos);
+
 	if (n != PAGE_SIZE)
 		return 0;
 	file->f_pos = pos;
diff --git a/fs/splice.c b/fs/splice.c
index 0878b852b355..2051700cda79 100644
--- a/fs/splice.c
+++ b/fs/splice.c
@@ -1234,8 +1234,8 @@ static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
  * as splice-from-memory, where the regular splice is splice-from-file (or
  * to file). In both cases the output is a pipe, naturally.
  */
-static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
-			     unsigned int flags)
+long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
+		      unsigned int flags)
 {
 	struct pipe_inode_info *pipe;
 	long ret = 0;
diff --git a/include/linux/splice.h b/include/linux/splice.h
index a55179fd60fc..0cd955cf5ee2 100644
--- a/include/linux/splice.h
+++ b/include/linux/splice.h
@@ -81,6 +81,8 @@ extern ssize_t splice_direct_to_actor(struct file *, struct splice_desc *,
 extern long do_splice(struct file *in, loff_t *off_in,
 		      struct file *out, loff_t *off_out,
 		      size_t len, unsigned int flags);
+extern long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
+			     unsigned int flags);
 
 extern long do_tee(struct file *in, struct file *out, size_t len,
 		   unsigned int flags);
-- 
2.20.1