php / php/php-src

Disable SHM reattachment

Open
#20,470 1 comment 2 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Feature Status: Needs Triage
Dominant language
C
Stars
40.4k
Forks
8.1k
Avg merge
2d 13h
Merged PRs (30d)
96

Description

Description

I'm opening this issue to restart the discussion started by @cmb69 in https://github.com/php/php-src/pull/17838 and https://github.com/php/php-src/pull/7865.

SHM reattachment leads to a number of problems, and likely fails often in practice, so I would like to disable it.

The reason we have SHM reattachment is that Windows doesn't support fork(), so multi-process SAPIs spawn independent processes. Therefore each process needs to reattach the Opcache SHM of previous processes.

This comes with a number of problems:

  1. The SHM must be attached at the same address in all processes. If the address is not free in one process, reattachment can not happen.
  2. When ASLR is enabled, some addresses referenced by SHM are invalid (such as opcode handler addresses stored in op arrays). Therefore, reattachment fails when ASLR is enabled (source).
  3. The memory layout of each process can differ regardless of ASLR, for example when DLLs were loaded in a different order. Therefore, SHM can not usually reference addresses out of SHM. This leads to various Windows-specific workarounds and special cases in the code base. For example, the inheritance cache is disabled for sub-classes of internal classes.

Additionally, this makes preloading incompatible with Windows, which blocks https://github.com/php/php-src/pull/18204.

In cases 1 and 2, the process fails with a fatal error, or fallsback to file_cache_only mode (when enabled), which is slower. So processes can randomly fail or be slower.

Case 2 implies that reattachment fails most of the times, now that ASLR is enabled by default.

Because of these, I believe that it would be beneficial to disable reattachment entirely.

Here is the impact on existing ways to deploy PHP, and what changes are required before we can disable reattachment:

IIS + FastCGI

In this configuration IIS itself manages PHP processes. Unfortunately this enforces a process-based model, and reattachment is required so that concurrent requests share the same cache.

I suggest that we recommend other configurations for performance-sensitive deployments. Disabling reattachment wouldn't impact development setups, so we can continue to support this configuration without reattachment.

Nginx + FastCGI

In this configuration, php-cgi manages child processes itself. Therefore we can switch this SAPI to threads in ZTS builds so that reattachment is not necessary anymore. This would likely make the SAPI more resource-efficient on Windows as well.

Apache + mod_php

Uses mpm_winnt, which spawns a single process and multiple threads.

Therefore, reattachment should not be necessary.

See https://github.com/php/php-src/pull/7865.

Additionally, @cmb69 started the implementation of an isapi-based SAPI:

IIS + isapi

Threaded, doesn't need reattachment

See https://github.com/php/php-src/pull/17838.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with ext/opcache/shared_alloc_win32.c and read the linked pull requests #17838 and #7865. Review how SHM reattachment affects IIS, Nginx, Apache, and the proposed ISAPI SAPI, then define the required deployment changes. Done means an agreed design and implementation that disables reattachment without breaking the supported threaded configurations.

Written by the indexing model from the issue text.

Assessment

Tech stack
c, php
Domain
backend, operating-systems
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.