Hi Greg, On 6/21/22 17:00, Greg Banks wrote:
Hi Michael and Alejandro,Did something useful happen to this patch? I don't see it in any public repo I'm aware of.
I think it's in the official git repository: <https://git.kernel.org/pub/scm/docs/man-pages/man-pages.git/commit/man5/nscd.conf.5?id=76954230a81187252201838fb329bff0beee4e8f> Is that correct? Thanks, Alex
Greg. ------------------------------------------------------------------------ *From:* Alejandro Colomar <alx.manpages@xxxxxxxxx> *Sent:* Friday, September 10, 2021 18:47 *To:* mtk.manpages@xxxxxxxxx <mtk.manpages@xxxxxxxxx>*Cc:* Greg Banks <gbanks@xxxxxxxxxxxx>; linux-man@xxxxxxxxxxxxxxx <linux-man@xxxxxxxxxxxxxxx>; Alejandro Colomar <alx.manpages@xxxxxxxxx>*Subject:* [PATCH 13/15] nscd.conf.5: describe reloading, clarifications From: Greg Banks <gbanks@xxxxxxxxxxxx> - Added a subsection of NOTES describing nscd's reloading behavior and providing advice on how to configure it. - Clarifications for the threads, reload-count, positive-time-to-live, check-files, and shared attributes. Derived by reading the nscd, libresolv and glibc source and some painful experience. Signed-off-by: Greg Banks <gbanks@xxxxxxxxxxxx> Signed-off-by: Alejandro Colomar <alx.manpages@xxxxxxxxx> --- man5/nscd.conf.5 | 115 +++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 112 insertions(+), 3 deletions(-) diff --git a/man5/nscd.conf.5 b/man5/nscd.conf.5 index 7356bf7c2..25ee4901b 100644 --- a/man5/nscd.conf.5 +++ b/man5/nscd.conf.5 @@ -1,5 +1,6 @@ .\" Copyright (c) 1999, 2000 SuSE GmbH Nuernberg, Germany .\" Author: Thorsten Kukuk <kukuk@xxxxxxx>+.\" Updates: Greg Banks <gbanks@xxxxxxxxxxxx> Copyright (c) 2021 Microsoft Corp..\" .\" %%%LICENSE_START(GPLv2+_SW_3_PARA) .\" This program is free software; you can redistribute it and/or @@ -53,9 +54,13 @@ The default is 0. .B threads .I number .RS -This is the number of threads that are started to wait for +This is the initial number of threads that are started to wait for requests. At least five threads will always be created. +The number of threads may increase dynamically up to +.B max\-threads +in response to demand from clients, +but never decreases. .RE .PP .B max\-threads @@ -83,9 +88,20 @@ Specifies the user who is allowed to request statistics. unlimited | .I number .RS-Limit on the number of times a cached entry gets reloaded without being used+Sets a limit on the number of times a cached entry +gets reloaded without being used before it gets removed. -The default is 5. +The limit can take values ranging from 0 to 254; +values 255 or higher behave the same as +.BR unlimited . +Limit values can be specified in either decimal +or hexadecimal with a "0x" prefix. +The special value +.B unlimited +is case-insensitive. +The default limit is 5. +A limit of 0 turns off the reloading feature. +See NOTES below for further discussion of reloading. .RE .PP .B paranoia @@ -128,6 +144,9 @@ in the specified cache for is in seconds. Larger values increase cache hit rates and reduce mean response times, but increase problems with cache coherence. +Note that for some name services (including specifically DNS) +the TTL returned from the name service is used and +this attribute is ignored. .RE .PP .B negative\-time\-to\-live @@ -166,6 +185,7 @@ The files are .IR /etc/passwd , .IR /etc/group , .IR /etc/hosts , +.IR /etc/resolv.conf , .IR /etc/services , and .IR /etc/netgroup . @@ -194,6 +214,8 @@ is shared with the clients so that they can directly search in them instead of having to ask the daemon over the socket each time a lookup is performed. The default is no. +Note that a cache miss will still result in +asking the daemon over the socket. .RE .PP .B max\-db\-size @@ -236,6 +258,93 @@ from the source code of and are used if not overridden in the configuration file. The default values used in the configuration file of your distribution might differ. +.SS Reloading +.BR nscd (8) +has a feature called reloading, +whose behavior can be surprising. +.PP +Reloading is enabled when the +.B reload-count +attribute has a non-zero value. +The default value in the source code enables reloading, +although your distribution may differ. +.PP +When reloading is enabled, +positive cached entries (the results of successful queries) +do not simply expire when their TTL is up. +Instead, at the expiry time, +.B nscd +will "reload", +i.e., +re-issue to the name service the same query that created the cached entry, +to get a new value to cache. +Depending on +.I /etc/nsswitch.conf +this may mean that a DNS, LDAP, or NIS request is made. +If the new query is successful, +reloading will repeat when the new value would expire, +until +.B reload-count +reloads have happened for the entry, +and only then will it actually be removed from the cache. +A request from a client which hits the entry will +reset the reload counter on the entry. +Purging the cache using +.I nscd\~-i +overrides the reload logic and removes the entry. +.PP +Reloading has the effect of extending cache entry TTLs +without compromising on cache coherency, +at the cost of additional load on the backing name service. +Whether this is a good idea on your system depends on +details of your applications' behavior, +your name service, +and the effective TTL values of your cache entries. +Note that for some name services +(for example, DNS), +the effective TTL is the value returned from the name service and +.I not +the value of the +.B positive\-time\-to\-live +attribute. +.PP +Please consider the following advice carefully: +.IP \(bu +If your application will make a second request for the same name, +after more than 1 TTL but before +.B reload\-count +TTLs, +and is sensitive to the latency of a cache miss, +then reloading may be a good idea for you. +.IP \(bu +If your name service is configured to return very short TTLs, +and your applications only make requests rarely under normal circumstances, +then reloading may result in additional load on your backing name service +without any benefit to applications, +which is probably a bad idea for you. +.IP \(bu +If your name service capacity is limited, +reloading may have the surprising effect of +increasing load on your name service instead of reducing it, +and may be a bad idea for you. +.IP \(bu +Setting +.B reload\-count +to +.B unlimited +is almost never a good idea, +as it will result in a cache that never expires entries +and puts never-ending additional load on the backing name service. +.PP +Some distributions have an init script for +.BR nscd (8) +with a +.I reload +command which uses +.I nscd\~-i +to purge the cache. +That use of the word "reload" is entirely different +from the "reloading" described here. .SH SEE ALSO .BR nscd (8) .\" .SH AUTHOR -- 2.33.0
-- Alejandro Colomar <http://www.alejandro-colomar.es/>
Attachment:
OpenPGP_signature
Description: OpenPGP digital signature
