[CRIU] [PATCH] pagemap: Support async pages reading by chunks

[Pavel Emelyanov]

When we collect all the read_page requests into one big preadv
call, the latter one may not read all the data in one go and
return less bytes read, than requested.

This is valid and already met in a bug :) So advance the iovec
set with the ret value and continue reading.

https://github.com/xemul/criu/issues/271

Signed-off-by: Pavel Emelyanov <xemul at virtuozzo.com>
---
 criu/pagemap.c | 54 ++++++++++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 48 insertions(+), 6 deletions(-)

diff --git a/criu/pagemap.c b/criu/pagemap.c
index a34bb2e..516e662 100644
--- a/criu/pagemap.c
+++ b/criu/pagemap.c
@@ -453,6 +453,31 @@ static void free_pagemaps(struct page_read *pr)
 	xfree(pr->pmes);
 }
 
+static void advance_piov(struct page_read_iov *piov, ssize_t len)
+{
+	ssize_t olen = len;
+	int onr = piov->nr;
+	piov->from += len;
+
+	while (len) {
+		struct iovec *cur = piov->to;
+
+		if (cur->iov_len <= len) {
+			piov->to++;
+			piov->nr--;
+			len -= cur->iov_len;
+			continue;
+		}
+
+		cur->iov_base += len;
+		cur->iov_len -= len;
+		break;
+	}
+
+	pr_info("Advanced iov %ld bytes, %d->%d iovs, %ld tail\n",
+			olen, onr, piov->nr, len);
+}
+
 static int process_async_reads(struct page_read *pr)
 {
 	int fd, ret = 0;
@@ -460,21 +485,38 @@ static int process_async_reads(struct page_read *pr)
 
 	fd = img_raw_fd(pr->pi);
 	list_for_each_entry_safe(piov, n, &pr->async, l) {
-		int ret;
-
+		ssize_t ret;
+		off_t start = piov->from;
+		struct iovec *iovs = piov->to;
+more:
 		ret = preadv(fd, piov->to, piov->nr, piov->from);
 		if (ret != piov->end - piov->from) {
-			pr_err("Can't read async pr bytes\n");
-			return -1;
+			if (ret < 0) {
+				pr_err("Can't read async pr bytes (%ld / %lu read, %lu off, %d iovs)\n",
+						ret, piov->end - piov->from, piov->from, piov->nr);
+				return -1;
+			}
+
+			/*
+			 * The preadv() can return less than requested. It's
+			 * valid and doesn't mean error or EOF. We should advance
+			 * the iovecs and continue
+			 *
+			 * Modify the piov in-place, we're going to drop this one
+			 * anyway.
+			 */
+
+			advance_piov(piov, ret);
+			goto more;
 		}
 
-		if (opts.auto_dedup && punch_hole(pr, piov->from, ret, false))
+		if (opts.auto_dedup && punch_hole(pr, start, ret, false))
 			return -1;
 
 		BUG_ON(pr->io_complete); /* FIXME -- implement once needed */
 
 		list_del(&piov->l);
-		xfree(piov->to);
+		xfree(iovs);
 		xfree(piov);
 	}
 
-- 
2.1.4