Skip to main content
How It Works5 min read

Bitcoin Core: A Paused Queue Can Still Fill Up

Two October 28 disclosures show different gaps in Bitcoin Core’s backpressure. Linux is exempt from the TCP urgent-data issue, but that exception does not extend to the paused-send issue.

Bitcoin Core: A Paused Queue Can Still Fill Up

AI-generated conceptual illustration: a stopped conveyor leaves a waiting hopper filling. It represents queue backpressure, not an actual Bitcoin node, documented attack or asset loss.

Key Points

  • Bitcoin Core disclosed CVE-2026-117350 and CVE-2026-117349 on October 28, rating both Medium. They describe two different ways peer-related memory could keep growing.
  • Linux is unaffected by the TCP urgent-data issue because it uses a different socket-monitoring path. The paused-send advisory has a different condition involving pre-BIP31 peers.
  • Both advisories name fixes in 29.5, 30.4, 31.2 and 32.1. One repair limits exceptional-event monitoring; the other adds a timeout for prolonged send pauses.
Inside this story

Chronological Timeline

2026-10-26Two fixes are merged

The advisories date the socket-monitoring and paused-send fixes to October 26 and record backports to branches 29–32.

2026-10-2731.2 and 32.1 are released

The disclosure timelines identify these as fixed releases on October 27.

2026-10-28Older-branch fixes and public disclosure

The timelines record 29.5 and 30.4 releases, followed by public disclosure of both issues, on October 28.

Two advisories need two sets of conditions

On October 28, 2026, Bitcoin Core published two memory-exhaustion advisories, CVE-2026-117350 and CVE-2026-117349. The maintainers classify both as Medium severity. They concern a node’s ability to keep running when a connected peer causes data to accumulate. Their shared lesson is about backpressure: slowing a producer or pausing a connection only limits memory if every path that can grow the waiting queue respects that limit.[1][2]

The platform qualification belongs to the first issue. The TCP urgent-data advisory explicitly excludes Linux and describes the affected Windows, macOS and BSD socket path. The second advisory concerns announcements accumulating for a send-paused peer that does not support BIP31’s nonced ping messages. The Linux exception from the first advisory cannot be carried over to the second. These are published vulnerability conditions; the disclosures do not establish that a particular service was attacked or lost funds.[1][2]

A receive pause bypassed by an exceptional event

Bitcoin Core limits how much data it queues for processing from each peer. When that waiting data becomes too large, it temporarily stops reading from the connection. On non-Linux platforms, the select() socket backend nevertheless kept watching the socket for exceptional conditions. Handling such an event could trigger another read, adding more data even though ordinary reads were paused. TCP urgent, or out-of-band, data could trigger that route.[1]

The advisory says this could drive unbounded memory growth and crash a connected node on the affected platforms. Linux uses the poll() backend and does not take this vulnerable path. PR #36665 changes when exceptional conditions are monitored: the socket must also be one from which the program is waiting to read. The PR explains that exceptional monitoring still has a purpose on Windows, where it helps detect failed connection attempts. The fix therefore makes the event path respect the read decision instead of assuming every socket event is harmless.[1][3]

A send pause without an effective lifetime limit

The other issue starts on the outgoing side. When a peer reads slowly and its outgoing message queue passes a limit, Bitcoin Core marks it as paused for sending. Some announcements can still be appended, especially frequent transaction inventory announcements, known as INVs. A pause flag does not mean that all producers have stopped adding work to that peer’s queue.[2]

Normally, a peer that fails to answer a nonced ping within 20 minutes is disconnected. The advisory explains that this disconnection was not enabled for peers lacking BIP31 support, so such a connection could remain paused and its queue could grow slowly without a bound. This is different from the first issue’s socket-event bypass. The advisory also says nodes running with -blocksonly are much less affected because transaction announcements account for most of the queued messages. That is a qualification about this second issue, not a claim of immunity or a substitute for the published fixes.[2]

One fix controls an event; the other bounds a pause

PR #36665 closes the mismatch between the receive-pause decision and exceptional socket monitoring. PR #36667 takes a different approach: it introduces a timeout for peers that stay paused for sending, using a 20-minute interval. Its stated goal is to stop retaining a connection whose send buffer cannot drain within a reasonable period, even when the usual ping-based disconnection does not help.[3][4]

The distinction matters when reading a security summary. A memory problem is not defined only by the amount of data in one packet. It can also arise from repeatedly admitting data around a stopped reader, or from retaining slowly accumulating work for too long. These repairs address different control points. Neither a platform exception for one path nor a lighter announcement workload establishes protection against every memory-exhaustion condition.[1][2][3][4]

Read the release dates and support status separately

Both advisories list 29.5, 30.4, 31.2 and 32.1 as releases containing the fixes. Their timelines place the merges and backports on October 26, releases 31.2 and 32.1 on October 27, and releases 29.5 and 30.4 plus public disclosure on October 28. The publication date is therefore not the date every fix first became available, and 32.1 is not the only fixed branch.[1][2]

A backported fix also does not promise continued maintenance. The 32.1 release notes state that 29.x and older versions have reached end of life. Readers evaluating these disclosures need the implementation and version, the operating-system path for the first issue, and the peer/queue conditions for the second. This article explains the official conditions and repairs; it does not reproduce an attack, benchmark memory growth or independently verify deployment across the Bitcoin network.[1][2][5]

The next question

Continue reading

Explore the topic through other cases and contexts.

Sources & References

  1. [1]Source 1: CVE-2026-117350 — TCP urgent data on non-Linux systemsBitcoin Core · 2026-10-28Accessed 2026-10-29
  2. [2]Source 2: CVE-2026-117349 — paused pre-BIP31 peersBitcoin Core · 2026-10-28Accessed 2026-10-29
  3. [3]Source 3: PR #36665 — align exception monitoring between socket backendsbitcoin/bitcoin · GitHubAccessed 2026-10-29
    Back to text123
  4. [4]Source 4: PR #36667 — enforce timeout for paused send peersbitcoin/bitcoin · GitHubAccessed 2026-10-29
    Back to text12
  5. [5]Source 5: Bitcoin Core 32.1 release notesBitcoin Core · 2026-10-27Accessed 2026-10-29
    Back to text1