perf(images): offload validate_url() DNS resolution with asyncio.to_thread (#25825)
validate_url() calls socket.getaddrinfo() for SSRF protection, which blocks the event loop for 100-700ms per DNS lookup. This affects: - Image generation (get_image_data) — every external image URL - Image editing (load_url_image) — every external image URL - OAuth profile pictures (_process_picture_url) — every login - Webhook delivery (post_webhook) — every notification - Image base64 conversion (get_image_base64_from_url) — chat images Wrap all 5 async call sites in asyncio.to_thread() so DNS resolution runs in the thread pool. The event loop remains free to serve other requests during the lookup. Benchmark (3 domains, 3 trials averaged): - BEFORE: max jitter 479ms, 1 blocked ping per trial - AFTER: max jitter 1ms, 0 blocked pings (324x improvement) Co-authored-by: Tim Baek <tim@openwebui.com>
This commit is contained in:
@@ -433,7 +433,7 @@ async def get_image_data(data: str, headers=None, trusted_base_url: str | None =
|
||||
if trusted_base_url and _is_same_origin(data, trusted_base_url):
|
||||
log.debug(f'Skipping URL validation for trusted backend: {data}')
|
||||
else:
|
||||
validate_url(data)
|
||||
await asyncio.to_thread(validate_url, data)
|
||||
session = await get_session()
|
||||
async with session.get(
|
||||
data,
|
||||
@@ -862,7 +862,7 @@ async def image_edits(
|
||||
# called only on the originally-submitted URL; following 3xx redirects
|
||||
# without re-validation would let an attacker reach private IPs via a
|
||||
# public host that redirects internally (e.g. cloud-metadata exfil).
|
||||
validate_url(data)
|
||||
await asyncio.to_thread(validate_url, data)
|
||||
# SSRF-safe session: re-checks the connect-time IP so a rebinding DNS answer
|
||||
# that passed validate_url cannot reach an internal address.
|
||||
async with get_ssrf_safe_session() as session:
|
||||
|
||||
Reference in New Issue
Block a user