{"payload":{"feedbackUrl":"https://github.com/orgs/community/discussions/53140","repo":{"id":251660789,"defaultBranch":"master","name":"zfs","ownerLogin":"KlaraSystems","currentUserCanPush":false,"isFork":true,"isEmpty":false,"createdAt":"2020-03-31T16:18:17.000Z","ownerAvatar":"https://avatars.githubusercontent.com/u/41072510?v=4","public":true,"private":false,"isOrgOwned":true},"refInfo":{"name":"","listCacheKey":"v0:1726078024.0","currentOid":""},"activityList":{"items":[{"before":"1713aa7b4d209616fab96a68e17a6fec6837247c","after":"a10e552b9992673626f7a2ffcc234337f23410c9","ref":"refs/heads/master","pushedAt":"2024-09-16T03:12:53.000Z","pushType":"push","commitsCount":1,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"Adding Direct IO Support\n\nAdding O_DIRECT support to ZFS to bypass the ARC for writes/reads.\r\n\r\nO_DIRECT support in ZFS will always ensure there is coherency between\r\nbuffered and O_DIRECT IO requests. This ensures that all IO requests,\r\nwhether buffered or direct, will see the same file contents at all\r\ntimes. Just as in other FS's , O_DIRECT does not imply O_SYNC. While\r\ndata is written directly to VDEV disks, metadata will not be synced\r\nuntil the associated TXG is synced.\r\nFor both O_DIRECT read and write request the offset and request sizes,\r\nat a minimum, must be PAGE_SIZE aligned. In the event they are not,\r\nthen EINVAL is returned unless the direct property is set to always (see\r\nbelow).\r\n\r\nFor O_DIRECT writes:\r\nThe request also must be block aligned (recordsize) or the write\r\nrequest will take the normal (buffered) write path. In the event that\r\nrequest is block aligned and a cached copy of the buffer in the ARC,\r\nthen it will be discarded from the ARC forcing all further reads to\r\nretrieve the data from disk.\r\n\r\nFor O_DIRECT reads:\r\nThe only alignment restrictions are PAGE_SIZE alignment. In the event\r\nthat the requested data is in buffered (in the ARC) it will just be\r\ncopied from the ARC into the user buffer.\r\n\r\nFor both O_DIRECT writes and reads the O_DIRECT flag will be ignored in\r\nthe event that file contents are mmap'ed. In this case, all requests\r\nthat are at least PAGE_SIZE aligned will just fall back to the buffered\r\npaths. If the request however is not PAGE_SIZE aligned, EINVAL will\r\nbe returned as always regardless if the file's contents are mmap'ed.\r\n\r\nSince O_DIRECT writes go through the normal ZIO pipeline, the\r\nfollowing operations are supported just as with normal buffered writes:\r\nChecksum\r\nCompression\r\nEncryption\r\nErasure Coding\r\nThere is one caveat for the data integrity of O_DIRECT writes that is\r\ndistinct for each of the OS's supported by ZFS.\r\nFreeBSD - FreeBSD is able to place user pages under write protection so\r\n any data in the user buffers and written directly down to the\r\n\t VDEV disks is guaranteed to not change. There is no concern\r\n\t with data integrity and O_DIRECT writes.\r\nLinux - Linux is not able to place anonymous user pages under write\r\n protection. Because of this, if the user decides to manipulate\r\n\tthe page contents while the write operation is occurring, data\r\n\tintegrity can not be guaranteed. However, there is a module\r\n\tparameter `zfs_vdev_direct_write_verify` that controls the\r\n\tif a O_DIRECT writes that can occur to a top-level VDEV before\r\n\ta checksum verify is run before the contents of the I/O buffer\r\n are committed to disk. In the event of a checksum verification\r\n\tfailure the write will return EIO. The number of O_DIRECT write\r\n\tchecksum verification errors can be observed by doing\r\n\t`zpool status -d`, which will list all verification errors that\r\n\thave occurred on a top-level VDEV. Along with `zpool status`, a\r\n\tZED event will be issues as `dio_verify` when a checksum\r\n\tverification error occurs.\r\n\r\nZVOLs and dedup is not currently supported with Direct I/O.\r\n\r\nA new dataset property `direct` has been added with the following 3\r\nallowable values:\r\ndisabled - Accepts O_DIRECT flag, but silently ignores it and treats\r\n\t the request as a buffered IO request.\r\nstandard - Follows the alignment restrictions outlined above for\r\n\t write/read IO requests when the O_DIRECT flag is used.\r\nalways - Treats every write/read IO request as though it passed\r\n O_DIRECT and will do O_DIRECT if the alignment restrictions\r\n\t are met otherwise will redirect through the ARC. This\r\n\t property will not allow a request to fail.\r\n\r\nThere is also a module parameter zfs_dio_enabled that can be used to\r\nforce all reads and writes through the ARC. By setting this module\r\nparameter to 0, it mimics as if the direct dataset property is set to\r\ndisabled.\r\n\r\nReviewed-by: Brian Behlendorf \r\nReviewed-by: Alexander Motin \r\nReviewed-by: Tony Hutter \r\nSigned-off-by: Brian Atkinson \r\nCo-authored-by: Mark Maybee \r\nCo-authored-by: Matt Macy \r\nCo-authored-by: Brian Behlendorf \r\nCloses #10018","shortMessageHtmlLink":"Adding Direct IO Support"}},{"before":"308f7c2f14deaa49b7b0d223564f48ccb6f5937b","after":"1713aa7b4d209616fab96a68e17a6fec6837247c","ref":"refs/heads/dedup_test_20240910","pushedAt":"2024-09-14T13:38:20.000Z","pushType":"push","commitsCount":1,"pusher":{"login":"allanjude","name":"Allan Jude","path":"/allanjude","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1096028?s=80&v=4"},"commit":{"message":"Remove set but not used variable in ddt.c (#16522)\n\nmodule/zfs/ddt.c:2612:6: error: variable 'total' set but not used\r\n\r\nSigned-off-by: Tino Reichardt \r\nReviewed-by: Alexander Motin \r\nReviewed-by: Tony Hutter ","shortMessageHtmlLink":"Remove set but not used variable in ddt.c (openzfs#16522)"}},{"before":"54888d8076f3cb7561b3118f425ed8829c61c0cc","after":"883cd30f908cb5c5a2df673ec85ebacac7933d11","ref":"refs/heads/parallel_dbuf_evict","pushedAt":"2024-09-12T11:59:12.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"0mp","name":"Mateusz Piotrowski","path":"/0mp","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/7978161?s=80&v=4"},"commit":{"message":"Implement parallel dbuf eviction\n\nIn the previous code, dbuf_evict_thread() would called dbuf_evict_one()\nin a look while dbuf_cache_above_lowater().\n\ndbuf_evict_one() would select a random sublist from the dbuf cache,\nthen walk it from the tail forward, attempting to acquire the lock on\neach object until it succeeded, then evict that object and return.\n\nAs the name suggests, it would evict only a single object from the\ncache. However, evicting one object is not likely to bring us below the\ndesired low water mark, so dbuf_evict_one() will be called again, where\nit will loop over all of the same busy objects again, until it founds\none it can evict.\n\nThis has been replaced with dbuf_evict_many() which takes a specific\nsublist as a parameter, as well as a desired amount of data to evict.\nIt then walks the sublist from the tail forward, evicting what it can\nuntil the number of bytes evicted satisfies the input parameter or\nthe head of the sublist is reached.\n\nThe dbuf_evict_thread now runs is parallel as well, allowing it to\nkeep up with demand more easily. For the dbuf cache, if the single\nthread was not able to keep up, ZFS would shift the work of evicting\nsome items to each incoming I/O thread. While that is still the case\nit should be seen much less often now that dbuf_evict is more efficient\nand no longer bottlenecked to a single thread.\n\nSponsored-by: Expensify, Inc.\nSponsored-by: Klara, Inc.\nCo-authored-by: Allan Jude \nCo-authored-by: Mateusz Piotrowski \nSigned-off-by: Alexander Stetsenko \nSigned-off-by: Allan Jude \nSigned-off-by: Mateusz Piotrowski ","shortMessageHtmlLink":"Implement parallel dbuf eviction"}},{"before":"d4a8574802b8836ac15daa6b7fa2d9a705b3d315","after":"54888d8076f3cb7561b3118f425ed8829c61c0cc","ref":"refs/heads/parallel_dbuf_evict","pushedAt":"2024-09-12T11:57:55.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"0mp","name":"Mateusz Piotrowski","path":"/0mp","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/7978161?s=80&v=4"},"commit":{"message":"Implement parallel dbuf eviction\n\nIn the previous code, dbuf_evict_thread() would called dbuf_evict_one()\nin a look while dbuf_cache_above_lowater().\n\ndbuf_evict_one() would select a random sublist from the dbuf cache,\nthen walk it from the tail forward, attempting to acquire the lock on\neach object until it succeeded, then evict that object and return.\n\nAs the name suggests, it would evict only a single object from the\ncache. However, evicting one object is not likely to bring us below the\ndesired low water mark, so dbuf_evict_one() will be called again, where\nit will loop over all of the same busy objects again, until it founds\none it can evict.\n\nThis has been replaced with dbuf_evict_many() which takes a specific\nsublist as a parameter, as well as a desired amount of data to evict.\nIt then walks the sublist from the tail forward, evicting what it can\nuntil the number of bytes evicted satisfies the input parameter or\nthe head of the sublist is reached.\n\nThe dbuf_evict_thread now runs is parallel as well, allowing it to\nkeep up with demand more easily. For the dbuf cache, if the single\nthread was not able to keep up, ZFS would shift the work of evicting\nsome items to each incoming I/O thread. While that is still the case\nit should be seen much less often now that dbuf_evict is more efficient\nand no longer bottlenecked to a single thread.\n\nSponsored-by: Expensify, Inc.\nSponsored-by: Klara, Inc.\nCo-authored-by: Allan Jude \nCo-authored-by: Mateusz Piotrowski \nSigned-off-by: Alexander Stetsenko \nSigned-off-by: Allan Jude \nSigned-off-by: Mateusz Piotrowski ","shortMessageHtmlLink":"Implement parallel dbuf eviction"}},{"before":"5b070d14972e888c336d6bb52edaacf9b7e75123","after":"d4a8574802b8836ac15daa6b7fa2d9a705b3d315","ref":"refs/heads/parallel_dbuf_evict","pushedAt":"2024-09-12T10:04:20.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"0mp","name":"Mateusz Piotrowski","path":"/0mp","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/7978161?s=80&v=4"},"commit":{"message":"Implement parallel dbuf eviction\n\nIn the previous code, dbuf_evict_thread() would called dbuf_evict_one()\nin a look while dbuf_cache_above_lowater().\n\ndbuf_evict_one() would select a random sublist from the dbuf cache,\nthen walk it from the tail forward, attempting to acquire the lock on\neach object until it succeeded, then evict that object and return.\n\nAs the name suggests, it would evict only a single object from the\ncache. However, evicting one object is not likely to bring us below the\ndesired low water mark, so dbuf_evict_one() will be called again, where\nit will loop over all of the same busy objects again, until it founds\none it can evict.\n\nThis has been replaced with dbuf_evict_many() which takes a specific\nsublist as a parameter, as well as a desired amount of data to evict.\nIt then walks the sublist from the tail forward, evicting what it can\nuntil the number of bytes evicted satisfies the input parameter or\nthe head of the sublist is reached.\n\nThe dbuf_evict_thread now runs is parallel as well, allowing it to\nkeep up with demand more easily. For the dbuf cache, if the single\nthread was not able to keep up, ZFS would shift the work of evicting\nsome items to each incoming I/O thread. While that is still the case\nit should be seen much less often now that dbuf_evict is more efficient\nand no longer bottlenecked to a single thread.\n\nSponsored-by: Expensify, Inc.\nSponsored-by: Klara, Inc.\nCo-authored-by: Allan Jude \nCo-authored-by: Mateusz Piotrowski \nSigned-off-by: Alexander Stetsenko \nSigned-off-by: Allan Jude \nSigned-off-by: Mateusz Piotrowski ","shortMessageHtmlLink":"Implement parallel dbuf eviction"}},{"before":"146fe45a44b644fe654317206fba3b20a13a2d19","after":"e1280262b5c4612dc0407b3d50e4c07ca1ecc29e","ref":"refs/heads/parallel_arc_evict","pushedAt":"2024-09-12T10:03:40.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"0mp","name":"Mateusz Piotrowski","path":"/0mp","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/7978161?s=80&v=4"},"commit":{"message":"Implement parallel ARC eviction\n\nRead and write performance can become limited by the arc_evict\nprocess being single threaded. Additional data cannot be added\nto the ARC until sufficient existing data is evicted.\n\nOn many-core systems with TBs of RAM, a single thread becomes\na significant bottleneck.\n\nWith the change we see a 25% increase in read and write throughput\n\nSponsored-by: Expensify, Inc.\nSponsored-by: Klara, Inc.\nCo-authored-by: Allan Jude \nCo-authored-by: Mateusz Piotrowski \nSigned-off-by: Alexander Stetsenko \nSigned-off-by: Allan Jude \nSigned-off-by: Mateusz Piotrowski ","shortMessageHtmlLink":"Implement parallel ARC eviction"}},{"before":null,"after":"308f7c2f14deaa49b7b0d223564f48ccb6f5937b","ref":"refs/heads/dedup_test_20240910","pushedAt":"2024-09-11T18:07:04.000Z","pushType":"branch_creation","commitsCount":0,"pusher":{"login":"allanjude","name":"Allan Jude","path":"/allanjude","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1096028?s=80&v=4"},"commit":{"message":"Fix an uninitialized data access (#16511)\n\nzfs_acl_node_alloc allocates an uninitialized data buffer, but upstack\r\nzfs_acl_chmod only partially initializes it. KMSAN reported that this\r\nmemory remained uninitialized at the point when it was read by\r\nlzjb_compress, which suggests a possible kernel memory disclosure bug.\r\n\r\nThe full KMSAN warning may be found in the PR.\r\nhttps://github.com/openzfs/zfs/pull/16511\r\n\r\nSigned-off-by:\tAlan Somers \r\nSponsored by:\tAxcient\r\nReviewed-by: Alexander Motin \r\nReviewed-by: Tony Hutter ","shortMessageHtmlLink":"Fix an uninitialized data access (openzfs#16511)"}},{"before":"b109925820fb79db3e37670c159977f03edd950f","after":"1713aa7b4d209616fab96a68e17a6fec6837247c","ref":"refs/heads/master","pushedAt":"2024-09-11T18:06:07.000Z","pushType":"push","commitsCount":6,"pusher":{"login":"allanjude","name":"Allan Jude","path":"/allanjude","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1096028?s=80&v=4"},"commit":{"message":"Remove set but not used variable in ddt.c (#16522)\n\nmodule/zfs/ddt.c:2612:6: error: variable 'total' set but not used\r\n\r\nSigned-off-by: Tino Reichardt \r\nReviewed-by: Alexander Motin \r\nReviewed-by: Tony Hutter ","shortMessageHtmlLink":"Remove set but not used variable in ddt.c (openzfs#16522)"}},{"before":"3dc96e07df08c9d54faf184d66ea1816ff8e3344","after":"dbee1a907f51064f9c02fd4f434352d17ae0864d","ref":"refs/heads/slack-compression","pushedAt":"2024-09-11T11:18:52.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"compress: add \"slack\" compression option\n\nThe \"slack\" option simply searches from the end of the block backwards\nto the last non-zero byte, and sets that position as the \"compressed\"\nsize.\n\nThis patch is highly experimental; please see the associated PR for\ndiscussion.\n\nSigned-off-by: Rob Norris \nSponsored-by: Klara, Inc.\nSponsored-by: Wasabi Technology, Inc.","shortMessageHtmlLink":"compress: add \"slack\" compression option"}},{"before":"2420ee6e12cb4bc4918fc88d44d59b486b86e58b","after":"b109925820fb79db3e37670c159977f03edd950f","ref":"refs/heads/master","pushedAt":"2024-09-07T01:53:56.000Z","pushType":"push","commitsCount":11,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"spa_prop_get: require caller to supply output nvlist\n\nAll callers to spa_prop_get() and spa_prop_get_nvlist() supplied their\r\nown preallocated nvlist (except ztest), so we can remove the option to\r\nhave them allocate one if none is supplied.\r\n\r\nThis sidesteps a bug in spa_prop_get(), where the error var wasn't\r\ninitialised, which could lead to the provided nvlist being freed at the\r\nend.\r\n\r\nSponsored-by: Klara, Inc.\r\nSponsored-by: Wasabi Technology, Inc.\r\nReviewed-by: Jorgen Lundman \r\nReviewed-by: Brian Behlendorf \r\nReviewed-by: Alexander Motin \r\nSigned-off-by: Rob Norris \r\nCloses #16505","shortMessageHtmlLink":"spa_prop_get: require caller to supply output nvlist"}},{"before":"3570e893e6ae22186a90316d1dcb4d48bedf0411","after":"5b070d14972e888c336d6bb52edaacf9b7e75123","ref":"refs/heads/parallel_dbuf_evict","pushedAt":"2024-09-04T23:17:37.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"allanjude","name":"Allan Jude","path":"/allanjude","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1096028?s=80&v=4"},"commit":{"message":"Implement parallel dbuf eviction\n\nIn the previous code, dbuf_evict_thread() would called dbuf_evict_one()\nin a look while dbuf_cache_above_lowater().\n\ndbuf_evict_one() would select a random sublist from the dbuf cache,\nthen walk it from the tail forward, attempting to acquire the lock on\neach object until it succeeded, then evict that object and return.\n\nAs the name suggests, it would evict only a single object from the\ncache. However, evicting one object is not likely to bring us below the\ndesired low water mark, so dbuf_evict_one() will be called again, where\nit will loop over all of the same busy objects again, until it founds\none it can evict.\n\nThis has been replaced with dbuf_evict_many() which takes a specific\nsublist as a parameter, as well as a desired amount of data to evict.\nIt then walks the sublist from the tail forward, evicting what it can\nuntil the number of bytes evicted satisfies the input parameter or\nthe head of the sublist is reached.\n\nThe dbuf_evict_thread now runs is parallel as well, allowing it to\nkeep up with demand more easily. For the dbuf cache, if the single\nthread was not able to keep up, ZFS would shift the work of evicting\nsome items to each incoming I/O thread. While that is still the case\nit should be seen much less often now that dbuf_evict is more efficient\nand no longer bottlenecked to a single thread.\n\nSponsored-by: Expensify, Inc.\nSponsored-by: Klara, Inc.\nCo-authored-by: Allan Jude \nSigned-off-by: Alexander Stetsenko \nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Implement parallel dbuf eviction"}},{"before":"f45bf2e957a0f535394fe157d216393881dba9e8","after":"146fe45a44b644fe654317206fba3b20a13a2d19","ref":"refs/heads/parallel_arc_evict","pushedAt":"2024-09-04T23:16:53.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"allanjude","name":"Allan Jude","path":"/allanjude","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1096028?s=80&v=4"},"commit":{"message":"Implement parallel ARC eviction\n\nRead and write performance can become limited by the arc_evict\nprocess being single threaded. Additional data cannot be added\nto the ARC until sufficient existing data is evicted.\n\nOn many-core systems with TBs of RAM, a single thread becomes\na significant bottleneck.\n\nWith the change we see a 25% increase in read and write throughput\n\nSponsored-by: Expensify, Inc.\nSponsored-by: Klara, Inc.\nCo-authored-by: Allan Jude \nSigned-off-by: Alexander Stetsenko \nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Implement parallel ARC eviction"}},{"before":"d3aad0f2eaa8c3d330797967e9f09b0636a19c40","after":"3570e893e6ae22186a90316d1dcb4d48bedf0411","ref":"refs/heads/parallel_dbuf_evict","pushedAt":"2024-08-29T18:36:28.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"allanjude","name":"Allan Jude","path":"/allanjude","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1096028?s=80&v=4"},"commit":{"message":"Implement parallel dbuf eviction\n\nIn the previous code, dbuf_evict_thread() would called dbuf_evict_one()\nin a look while dbuf_cache_above_lowater().\n\ndbuf_evict_one() would select a random sublist from the dbuf cache,\nthen walk it from the tail forward, attempting to acquire the lock on\neach object until it succeeded, then evict that object and return.\n\nAs the name suggests, it would evict only a single object from the\ncache. However, evicting one object is not likely to bring us below the\ndesired low water mark, so dbuf_evict_one() will be called again, where\nit will loop over all of the same busy objects again, until it founds\none it can evict.\n\nThis has been replaced with dbuf_evict_many() which takes a specific\nsublist as a parameter, as well as a desired amount of data to evict.\nIt then walks the sublist from the tail forward, evicting what it can\nuntil the number of bytes evicted satisfies the input parameter or\nthe head of the sublist is reached.\n\nThe dbuf_evict_thread now runs is parallel as well, allowing it to\nkeep up with demand more easily. For the dbuf cache, if the single\nthread was not able to keep up, ZFS would shift the work of evicting\nsome items to each incoming I/O thread. While that is still the case\nit should be seen much less often now that dbuf_evict is more efficient\nand no longer bottlenecked to a single thread.\n\nSponsored-by: Expensify, Inc.\nSponsored-by: Klara, Inc.\nCo-authored-by: Allan Jude \nSigned-off-by: Alexander Stetsenko \nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Implement parallel dbuf eviction"}},{"before":"4cd510d8485594ba50a2c958e954f0baa7e53e74","after":"f45bf2e957a0f535394fe157d216393881dba9e8","ref":"refs/heads/parallel_arc_evict","pushedAt":"2024-08-29T18:19:48.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"allanjude","name":"Allan Jude","path":"/allanjude","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1096028?s=80&v=4"},"commit":{"message":"Implement parallel ARC eviction\n\nRead and write performance can become limited by the arc_evict\nprocess being single threaded. Additional data cannot be added\nto the ARC until sufficient existing data is evicted.\n\nOn many-core systems with TBs of RAM, a single thread becomes\na significant bottleneck.\n\nWith the change we see a 25% increase in read and write throughput\n\nSponsored-by: Expensify, Inc.\nSponsored-by: Klara, Inc.\nCo-authored-by: Allan Jude \nSigned-off-by: Alexander Stetsenko \nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Implement parallel ARC eviction"}},{"before":null,"after":"d3aad0f2eaa8c3d330797967e9f09b0636a19c40","ref":"refs/heads/parallel_dbuf_evict","pushedAt":"2024-08-29T16:46:27.000Z","pushType":"branch_creation","commitsCount":0,"pusher":{"login":"allanjude","name":"Allan Jude","path":"/allanjude","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1096028?s=80&v=4"},"commit":{"message":"Implement parallel dbuf eviction\n\nIn the previous code, dbuf_evict_thread() would called dbuf_evict_one()\nin a look while dbuf_cache_above_lowater().\n\ndbuf_evict_one() would select a random sublist from the dbuf cache,\nthen walk it from the tail forward, attempting to acquire the lock on\neach object until it succeeded, then evict that object and return.\n\nAs the name suggests, it would evict only a single object from the cache.\nHowever, evicting one object is not likely to bring us below the\ndesired low water mark, so dbuf_evict_one() will be called again, where\nit will loop over all of the same busy objects again, until it founds\none it can evict.\n\nThis has been replaced with dbuf_evict_many() which takes a specific\nsublist as a parameter, as well as a desired amount of data to evict.\nIt then walks the sublist from the tail forward, evicting what it can\nuntil the number of bytes evicted satisfies the input parameter or\nthe head of the sublist is reached.\n\nThe dbuf_evict_thread now runs is parallel as well, allowing it to\nkeep up with demand more easily. For the dbuf cache, if the single\nthread was not able to keep up, ZFS would shift the work of evicting\nsome items to each incoming I/O thread. While that is still the case\nit should be seen much less often now that dbuf_evict is more efficient\nand no longer bottlenecked to a single thread.\n\nSponsored-by: Expensify, Inc.\nSponsored-by: Klara, Inc.\nCo-authored-by: Allan Jude \nSigned-off-by: Alexander Stetsenko \nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Implement parallel dbuf eviction"}},{"before":null,"after":"4cd510d8485594ba50a2c958e954f0baa7e53e74","ref":"refs/heads/parallel_arc_evict","pushedAt":"2024-08-29T16:39:13.000Z","pushType":"branch_creation","commitsCount":0,"pusher":{"login":"allanjude","name":"Allan Jude","path":"/allanjude","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1096028?s=80&v=4"},"commit":{"message":"Implement parallel ARC eviction\n\nRead and write performance can become limited by the arc_evict\nprocess being single threaded. Additional data cannot be added\nto the ARC until sufficient existing data is evicted.\n\nOn many-core systems with TBs of RAM, a single thread becomes\na significant bottleneck.\n\nWith the change we see a 25% increase in read and write throughput\n\nSponsored-by: Expensify, Inc.\nSponsored-by: Klara, Inc.\nCo-authored-by: Allan Jude \nSigned-off-by: Alexander Stetsenko \nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Implement parallel ARC eviction"}},{"before":"5b542d2ad9bee142a485fee0af0c938775c1c69b","after":null,"ref":"refs/heads/zpool_reguid","pushedAt":"2024-08-26T16:30:12.000Z","pushType":"branch_deletion","commitsCount":0,"pusher":{"login":"0mp","name":"Mateusz Piotrowski","path":"/0mp","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/7978161?s=80&v=4"}},{"before":"77a797a3823c12268e7f1d73f5e024a77b2f582a","after":"2420ee6e12cb4bc4918fc88d44d59b486b86e58b","ref":"refs/heads/master","pushedAt":"2024-08-26T12:02:10.000Z","pushType":"push","commitsCount":27,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"spl-taskq: fix task counts for delayed and cancelled tasks\n\nDispatched delayed tasks were not added to tasks_total, and cancelled\r\ntasks were not removed. This notably could make tasks_total go to\r\nUNIT64_MAX, but just generally meant the count could be wrong. So lets\r\nnot!\r\n\r\nSponsored-by: Klara, Inc.\r\nSponsored-by: Syneto\r\nReviewed-by: Brian Behlendorf \r\nSigned-off-by: Rob Norris \r\nCloses #16473","shortMessageHtmlLink":"spl-taskq: fix task counts for delayed and cancelled tasks"}},{"before":"3d480a164f18497c7a450d87f95c99fba3117c75","after":"5b542d2ad9bee142a485fee0af0c938775c1c69b","ref":"refs/heads/zpool_reguid","pushedAt":"2024-08-22T21:56:54.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"0mp","name":"Mateusz Piotrowski","path":"/0mp","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/7978161?s=80&v=4"},"commit":{"message":"zpool: Provide GUID to zpool-reguid(8) with -g\n\nThis commit extends the zpool-reguid(8) command with a -g flag, which\nallows the user to specify the GUID to set.\n\nThis change also adds some general tests for zpool-reguid(8).\n\nSponsored-by: Wasabi Technology, Inc.\nSponsored-by: Klara, Inc.\nSigned-off-by: Mateusz Piotrowski <0mp@FreeBSD.org>","shortMessageHtmlLink":"zpool: Provide GUID to zpool-reguid(8) with -g"}},{"before":"0e58adde947a262ad584b5a80e4a0a61800b9a5c","after":"e8342dab9340b3cc1c1319c517cdf27ce64405a2","ref":"refs/heads/fdt-rel-prune-pr","pushedAt":"2024-08-20T14:50:41.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"allanjude","name":"Allan Jude","path":"/allanjude","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1096028?s=80&v=4"},"commit":{"message":"Address review feedback (to be squashed)\n\nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Address review feedback (to be squashed)"}},{"before":"bdf4d6be1de870b16d4f7997b235d9f19dd7e30e","after":"77a797a3823c12268e7f1d73f5e024a77b2f582a","ref":"refs/heads/master","pushedAt":"2024-08-18T01:51:29.000Z","pushType":"push","commitsCount":15,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"Enable L2 cache of all (MRU+MFU) metadata but MFU data only\n\n`l2arc_mfuonly` was added to avoid wasting L2 ARC on read-once MRU\r\ndata and metadata. However it can be useful to cache as much\r\nmetadata as possible while, at the same time, restricting data\r\ncache to MFU buffers only.\r\n\r\nThis patch allow for such behavior by setting `l2arc_mfuonly` to 2\r\n(or higher). The list of possible values is the following:\r\n0: cache both MRU and MFU for both data and metadata;\r\n1: cache only MFU for both data and metadata;\r\n2: cache both MRU and MFU for metadata, but only MFU for data.\r\n\r\nReviewed-by: Alexander Motin \r\nReviewed-by: Brian Behlendorf \r\nSigned-off-by: Gionatan Danti \r\nCloses #16343 \r\nCloses #16402","shortMessageHtmlLink":"Enable L2 cache of all (MRU+MFU) metadata but MFU data only"}},{"before":null,"after":"0e58adde947a262ad584b5a80e4a0a61800b9a5c","ref":"refs/heads/fdt-rel-prune-pr","pushedAt":"2024-08-16T13:02:19.000Z","pushType":"branch_creation","commitsCount":0,"pusher":{"login":"allanjude","name":"Allan Jude","path":"/allanjude","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1096028?s=80&v=4"},"commit":{"message":"Address review feedback (to be squashed)\n\nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Address review feedback (to be squashed)"}},{"before":"b65f86ad7bebacd70a1d69ecab4bd3f10821c766","after":"bc877feb397e34c89580964af53ec2e50eb12cc9","ref":"refs/heads/fdt-rel-log","pushedAt":"2024-08-16T05:31:39.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"Man page updates for dmu_ddt_copies\n\nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Man page updates for dmu_ddt_copies"}},{"before":"6c82951d111bb4c8a426e5f58a87ac80a4996fc1","after":"bdf4d6be1de870b16d4f7997b235d9f19dd7e30e","ref":"refs/heads/master","pushedAt":"2024-08-16T00:01:20.000Z","pushType":"push","commitsCount":51,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"linux/zvol_os: fix zvol queue limits initialization\n\nzvol queue limits initialization depends on `zv_volblocksize`, but it is\r\ninitialized later, leading to several limits being initialized with\r\nincorrect values, including `max_discard_*` limits. This also causes\r\n`blkdiscard` command to consistently fail, as `blk_ioctl_discard` reads\r\n`bdev_max_discard_sectors()` limits as 0, leading to failure. The fix is\r\nstraightforward: initialize `zv->zv_volblocksize` early, before setting\r\nthe queue limits. This PR should fix `zvol/zvol_misc/zvol_misc_trim`\r\nfailure on recent PRs, as the test case issues `blkdiscard` for a zvol.\r\nAdditionally, `zvol_misc_trim` was recently enabled in `6c7d41a`,\r\nwhich is why the issue wasn't identified earlier.\r\n\r\nReviewed-by: Brian Behlendorf \r\nReviewed-by: Tony Hutter \r\nSigned-off-by: Ameer Hamza \r\nCloses #16454","shortMessageHtmlLink":"linux/zvol_os: fix zvol queue limits initialization"}},{"before":"c3930e26b684a3765d2ab4c73ca3ac1cb092aec4","after":"b65f86ad7bebacd70a1d69ecab4bd3f10821c766","ref":"refs/heads/fdt-rel-log","pushedAt":"2024-08-16T00:00:12.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"Man page updates for dmu_ddt_copies\n\nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Man page updates for dmu_ddt_copies"}},{"before":"4eb00a71d10c9939a5c9a48f8ecec8018a93e6ca","after":"c61047c0e51c197d978a4682e1c1ed91cc255dde","ref":"refs/heads/fdt-rel-feature","pushedAt":"2024-08-16T00:00:12.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"zdb: rework DDT block count and leak check to just count the blocks\n\nThe upcoming dedup features break the long held assumption that all\nblocks on disk with a 'D' dedup bit will always be present in the DDT,\nor will have the same set of DVA allocations on disk as in the DDT.\n\nIf the DDT is no longer a complete picture of all the dedup blocks that\nwill be and should be on disk, then it does us no good to walk and prime\nit up front, since it won't necessarily match up with every block we'll\nsee anyway.\n\nInstead, we rework things here to be more like the BRT checks. When we\nsee a dedup'd block, we look it up in the DDT, consume a refcount, and\nfor the second-or-later instances, count them as duplicates.\n\nThe DDT and BRT are moved ahead of the space accounting. This will\nbecome important for the \"flat\" feature, which may need to count a\nmodified version of the block.\n\nCo-authored-by: Allan Jude \nCo-authored-by: Don Brady \nSigned-off-by: Rob Norris \nSponsored-by: Klara, Inc.\nSponsored-by: iXsystems, Inc.","shortMessageHtmlLink":"zdb: rework DDT block count and leak check to just count the blocks"}},{"before":"e119f942b5d6bbe05978874152ac110685332e33","after":"eb0cb798ba50a5e9a2b2f0fd0ef9bd5d037cc571","ref":"refs/heads/fdt-rel-flat","pushedAt":"2024-08-16T00:00:12.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"ddt: add \"flat phys\" feature\n\nTraditional dedup keeps a separate ddt_phys_t \"type\" for each possible\ncount of DVAs (that is, copies=) parameter. Each of these are tracked\nindependently of each other, and have their own set of DVAs. This leads\nto an (admittedly rare) situation where you can create as many as six\ncopies of the data, by changing the copies= parameter between copying.\nThis is both a waste of storage on disk, but also a waste of space in\nthe stored DDT entries, since there never needs to be more than three\nDVAs to handle all possible values of copies=.\n\nThis commit adds a new FDT feature, DDT_FLAG_FLAT. When active, only the\nfirst ddt_phys_t is used. Each time a block is written with the dedup\nbit set, this single phys is checked to see if it has enough DVAs to\nfulfill the request. If it does, the block is filled with the saved DVAs\nas normal. If not, an adjusted write is issued to create as many extra\ncopies as are needed to fulfill the request, which are then saved into\nthe entry too.\n\nBecause a single phys is no longer an all-or-nothing, but can be\ntransitioning from fewer to more DVAs, the write path now has to keep a\ncopy of the previous \"known good\" DVA set so we can revert to it in case\nan error occurs. zio_ddt_write() has been restructured and heavily\ncommented to make it much easier to see what's happening.\n\nBackwards compatibility is maintained simply by allocating four\nddt_phys_t when the DDT_FLAG_FLAT flag is not set, and updating the phys\nselection macros to check the flag. In the old arrangement, each number\nof copies gets a whole phys, so it will always have either zero or all\nnecessary DVAs filled, with no in-between, so the old behaviour\nnaturally falls out of the new code.\n\nSigned-off-by: Rob Norris \nCo-authored-by: Don Brady \nSponsored-by: Klara, Inc.\nSponsored-by: iXsystems, Inc.","shortMessageHtmlLink":"ddt: add \"flat phys\" feature"}},{"before":"4652962af2099cd7e501ff96a9732c49d71038f6","after":"c3930e26b684a3765d2ab4c73ca3ac1cb092aec4","ref":"refs/heads/fdt-rel-log","pushedAt":"2024-08-13T20:18:37.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"Man page updates for dmu_ddt_copies\n\nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Man page updates for dmu_ddt_copies"}},{"before":"e8fa0c827650bcb8c258d9f70972c10c95e8391d","after":"4652962af2099cd7e501ff96a9732c49d71038f6","ref":"refs/heads/fdt-rel-log","pushedAt":"2024-08-13T11:58:09.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"Man page updates for dmu_ddt_copies\n\nSigned-off-by: Allan Jude ","shortMessageHtmlLink":"Man page updates for dmu_ddt_copies"}},{"before":"2a46268b6649b8fe93f07fd30d209f0d68b860ce","after":"e119f942b5d6bbe05978874152ac110685332e33","ref":"refs/heads/fdt-rel-flat","pushedAt":"2024-08-13T11:56:06.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"robn","name":"Rob Norris","path":"/robn","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/130670?s=80&v=4"},"commit":{"message":"ddt: add \"flat phys\" feature\n\nTraditional dedup keeps a separate ddt_phys_t \"type\" for each possible\ncount of DVAs (that is, copies=) parameter. Each of these are tracked\nindependently of each other, and have their own set of DVAs. This leads\nto an (admittedly rare) situation where you can create as many as six\ncopies of the data, by changing the copies= parameter between copying.\nThis is both a waste of storage on disk, but also a waste of space in\nthe stored DDT entries, since there never needs to be more than three\nDVAs to handle all possible values of copies=.\n\nThis commit adds a new FDT feature, DDT_FLAG_FLAT. When active, only the\nfirst ddt_phys_t is used. Each time a block is written with the dedup\nbit set, this single phys is checked to see if it has enough DVAs to\nfulfill the request. If it does, the block is filled with the saved DVAs\nas normal. If not, an adjusted write is issued to create as many extra\ncopies as are needed to fulfill the request, which are then saved into\nthe entry too.\n\nBecause a single phys is no longer an all-or-nothing, but can be\ntransitioning from fewer to more DVAs, the write path now has to keep a\ncopy of the previous \"known good\" DVA set so we can revert to it in case\nan error occurs. zio_ddt_write() has been restructured and heavily\ncommented to make it much easier to see what's happening.\n\nBackwards compatibility is maintained simply by allocating four\nddt_phys_t when the DDT_FLAG_FLAT flag is not set, and updating the phys\nselection macros to check the flag. In the old arrangement, each number\nof copies gets a whole phys, so it will always have either zero or all\nnecessary DVAs filled, with no in-between, so the old behaviour\nnaturally falls out of the new code.\n\nSigned-off-by: Rob Norris \nCo-authored-by: Don Brady \nSponsored-by: Klara, Inc.\nSponsored-by: iXsystems, Inc.","shortMessageHtmlLink":"ddt: add \"flat phys\" feature"}}],"hasNextPage":true,"hasPreviousPage":false,"activityType":"all","actor":null,"timePeriod":"all","sort":"DESC","perPage":30,"cursor":"Y3Vyc29yOnYyOpK7MjAyNC0wOS0xNlQwMzoxMjo1My4wMDAwMDBazwAAAAS2_3gF","startCursor":"Y3Vyc29yOnYyOpK7MjAyNC0wOS0xNlQwMzoxMjo1My4wMDAwMDBazwAAAAS2_3gF","endCursor":"Y3Vyc29yOnYyOpK7MjAyNC0wOC0xM1QxMTo1NjowNi4wMDAwMDBazwAAAASZIMDD"}},"title":"Activity ยท KlaraSystems/zfs"}