unix,readv: always permit partial reads to return

For simplicity and predictability (since the user must handle the retry
anyways), always emit exactly one readv/pread/preadv syscall and return
that result to the user.

By contrast, write needs to preserve order, so it needs to keep retrying
the operation until it finishes before retiring the req from the queue.

Fixes: https://github.com/nodejs/node/issues/16601
PR-URL: https://github.com/libuv/libuv/pull/1742
Refs: https://github.com/libuv/libuv/pull/640
Refs: https://github.com/libuv/libuv/issues/1720
Reviewed-By: Santiago Gimeno <santiago.gimeno@gmail.com>

1 files changed, 25 insertions, 10 deletions

diff --git a/src/unix/fs.c b/src/unix/fs.c
index 8d773aa1..35bcc6f6 100644
--- a/src/unix/fs.c
+++ b/src/unix/fs.c
@@ -262,17 +262,25 @@ static ssize_t uv__fs_read(uv_fs_t* req) {
 #if defined(__linux__)
   static int no_preadv;
 #endif
+  unsigned int iovmax;
   ssize_t result;
 
 #if defined(_AIX)
   struct stat buf;
-  if(fstat(req->file, &buf))
-    return -1;
-  if(S_ISDIR(buf.st_mode)) {
+  result = fstat(req->file, &buf);
+  if (result)
+    goto done;
+  if (S_ISDIR(buf.st_mode)) {
     errno = EISDIR;
-    return -1;
+    result -1;
+    goto done;
   }
 #endif /* defined(_AIX) */
+
+  iovmax = uv__getiovmax();
+  if (req->nbufs > iovmax)
+    req->nbufs = iovmax;
+
   if (req->off < 0) {
     if (req->nbufs == 1)
       result = read(req->file, req->bufs[0].base, req->bufs[0].len);
@@ -309,6 +317,13 @@ static ssize_t uv__fs_read(uv_fs_t* req) {
   }
 
 done:
+  /* Early cleanup of bufs allocation, since we're done with it. */
+  if (req->bufs != req->bufsml)
+    uv__free(req->bufs);
+
+  req->bufs = NULL;
+  req->nbufs = 0;
+
   return result;
 }
 
@@ -1023,8 +1038,7 @@ static size_t uv__fs_buf_offset(uv_buf_t* bufs, size_t size) {
   return offset;
 }
 
-typedef ssize_t (*uv__fs_buf_iter_processor)(uv_fs_t* req);
-static ssize_t uv__fs_buf_iter(uv_fs_t* req, uv__fs_buf_iter_processor process) {
+static ssize_t uv__fs_write_all(uv_fs_t* req) {
   unsigned int iovmax;
   unsigned int nbufs;
   uv_buf_t* bufs;
@@ -1042,7 +1056,7 @@ static ssize_t uv__fs_buf_iter(uv_fs_t* req, uv__fs_buf_iter_processor process)
       req->nbufs = iovmax;
 
     do
-      result = process(req);
+      result = uv__fs_write(req);
     while (result < 0 && errno == EINTR);
 
     if (result <= 0) {
@@ -1076,7 +1090,8 @@ static void uv__fs_work(struct uv__work* w) {
   ssize_t r;
 
   req = container_of(w, uv_fs_t, work_req);
-  retry_on_eintr = !(req->fs_type == UV_FS_CLOSE);
+  retry_on_eintr = !(req->fs_type == UV_FS_CLOSE ||
+                     req->fs_type == UV_FS_READ);
 
   do {
     errno = 0;
@@ -1105,7 +1120,7 @@ static void uv__fs_work(struct uv__work* w) {
     X(MKDIR, mkdir(req->path, req->mode));
     X(MKDTEMP, uv__fs_mkdtemp(req));
     X(OPEN, uv__fs_open(req));
-    X(READ, uv__fs_buf_iter(req, uv__fs_read));
+    X(READ, uv__fs_read(req));
     X(SCANDIR, uv__fs_scandir(req));
     X(READLINK, uv__fs_readlink(req));
     X(REALPATH, uv__fs_realpath(req));
@@ -1116,7 +1131,7 @@ static void uv__fs_work(struct uv__work* w) {
     X(SYMLINK, symlink(req->path, req->new_path));
     X(UNLINK, unlink(req->path));
     X(UTIME, uv__fs_utime(req));
-    X(WRITE, uv__fs_buf_iter(req, uv__fs_write));
+    X(WRITE, uv__fs_write_all(req));
     default: abort();
     }
 #undef X