练习5.13: 修改crawl,使其能保存发现的页面,必要时,可以创建目录来保存这些页面。只保存来自原始域名下的页面。假设初始页面在golang.org下,就不
要保存vimeo.com下的页面。
package main import ( "fmt" "io" "io/ioutil" "links" "log" "net/http" "net/url" "os" ) /* 练习5.13: 修改crawl,使其能保存发现的页面,必要时,可以创建目录来保存这些页面。只保存来自原始域名下的页面。假设初始页面在golang.org下,就不要保存vimeo.com下的页面。 */ var sum int func main() { breadthFirst(crawl, os.Args[1:]) } /* 抓取页面的所有连接 */ func crawl(url string) []string { sum++ go save(url) fmt.Printf("%d|%s\n", sum, url) list, err := links.Extract(url) if err != nil { log.Print(err) } return list } /* 保存页面到文件 */ func save(u string) bool { urlObj, _ := url.Parse(u) path := "/tmp/crawl/" + urlObj.Host if urlObj.Path == "" || urlObj.Path == "/" { urlObj.Path = "/index.html" } filename := path + urlObj.Path //重点注意文件名 fmt.Println(filename) //打开文件 f, _ := os.OpenFile(filename, os.O_CREATE|os.O_APPEND|os.O_RDWR, 0755) //读取链接 resp, geterr := http.Get(u) if geterr != nil || resp.StatusCode != http.StatusOK { //resp.Body.Close() return false } body, _ := ioutil.ReadAll(resp.Body) //fmt.Println(body) //创建保存目录 _, err := os.Stat(path) if err != nil { os.MkdirAll(path, 0755) } io.WriteString(f, string(body)) resp.Body.Close() body = nil return true } /* 广度优先算法 */ // breadthFirst calls f for each item in the worklist. // Any items returned by f are added to the worklist. // f is called at most once for each item. func breadthFirst(f func(item string) []string, worklist []string) { seen := make(map[string]bool) for len(worklist) > 0 { items := worklist worklist = nil for _, item := range items { if !seen[item] { seen[item] = true worklist = append(worklist, f(item)...) } } } }
原文地址:https://www.cnblogs.com/taoshihan/p/8875218.html
时间: 2024-10-03 19:17:09