LCOV - differential code coverage report
Current view: top level - src/backend/storage/aio - method_io_uring.c (source / functions) Coverage Total Hit UBC CBC
Current: c70b6db34ffeab48beef1fb4ce61bcad3772b8dd vs 06473f5a344df8c9594ead90a609b86f6724cff8 Lines: 74.1 % 193 143 50 143
Current Date: 2025-09-06 07:49:51 +0900 Functions: 100.0 % 12 12 12
Baseline: lcov-20250906-005545-baseline Branches: 44.6 % 112 50 62 50
Baseline Date: 2025-09-05 08:21:35 +0100 Line coverage date bins:
Legend: Lines:     hit not hit
Branches: + taken - not taken # not executed
(7,30] days: 50.0 % 2 1 1 1
(30,360] days: 74.3 % 191 142 49 142
Function coverage date bins:
(30,360] days: 100.0 % 12 12 12
Branch coverage date bins:
(7,30] days: 25.0 % 4 1 3 1
(30,360] days: 45.4 % 108 49 59 49

 Age         Owner                    Branch data    TLA  Line data    Source code
                                  1                 :                : /*-------------------------------------------------------------------------
                                  2                 :                :  *
                                  3                 :                :  * method_io_uring.c
                                  4                 :                :  *    AIO - perform AIO using Linux' io_uring
                                  5                 :                :  *
                                  6                 :                :  * For now we create one io_uring instance for each backend. These io_uring
                                  7                 :                :  * instances have to be created in postmaster, during startup, to allow other
                                  8                 :                :  * backends to process IO completions, if the issuing backend is currently
                                  9                 :                :  * busy doing other things. Other backends may not use another backend's
                                 10                 :                :  * io_uring instance to submit IO, that'd require additional locking that
                                 11                 :                :  * would likely be harmful for performance.
                                 12                 :                :  *
                                 13                 :                :  * We likely will want to introduce a backend-local io_uring instance in the
                                 14                 :                :  * future, e.g. for FE/BE network IO.
                                 15                 :                :  *
                                 16                 :                :  * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
                                 17                 :                :  * Portions Copyright (c) 1994, Regents of the University of California
                                 18                 :                :  *
                                 19                 :                :  * IDENTIFICATION
                                 20                 :                :  *    src/backend/storage/aio/method_io_uring.c
                                 21                 :                :  *
                                 22                 :                :  *-------------------------------------------------------------------------
                                 23                 :                :  */
                                 24                 :                : 
                                 25                 :                : #include "postgres.h"
                                 26                 :                : 
                                 27                 :                : /* included early, for IOMETHOD_IO_URING_ENABLED */
                                 28                 :                : #include "storage/aio.h"
                                 29                 :                : 
                                 30                 :                : #ifdef IOMETHOD_IO_URING_ENABLED
                                 31                 :                : 
                                 32                 :                : #include <sys/mman.h>
                                 33                 :                : #include <unistd.h>
                                 34                 :                : 
                                 35                 :                : #include <liburing.h>
                                 36                 :                : 
                                 37                 :                : #include "miscadmin.h"
                                 38                 :                : #include "storage/aio_internal.h"
                                 39                 :                : #include "storage/fd.h"
                                 40                 :                : #include "storage/proc.h"
                                 41                 :                : #include "storage/shmem.h"
                                 42                 :                : #include "storage/lwlock.h"
                                 43                 :                : #include "storage/procnumber.h"
                                 44                 :                : #include "utils/wait_event.h"
                                 45                 :                : 
                                 46                 :                : 
                                 47                 :                : /* number of completions processed at once */
                                 48                 :                : #define PGAIO_MAX_LOCAL_COMPLETED_IO 32
                                 49                 :                : 
                                 50                 :                : 
                                 51                 :                : /* Entry points for IoMethodOps. */
                                 52                 :                : static size_t pgaio_uring_shmem_size(void);
                                 53                 :                : static void pgaio_uring_shmem_init(bool first_time);
                                 54                 :                : static void pgaio_uring_init_backend(void);
                                 55                 :                : static int  pgaio_uring_submit(uint16 num_staged_ios, PgAioHandle **staged_ios);
                                 56                 :                : static void pgaio_uring_wait_one(PgAioHandle *ioh, uint64 ref_generation);
                                 57                 :                : 
                                 58                 :                : /* helper functions */
                                 59                 :                : static void pgaio_uring_sq_from_io(PgAioHandle *ioh, struct io_uring_sqe *sqe);
                                 60                 :                : 
                                 61                 :                : 
                                 62                 :                : const IoMethodOps pgaio_uring_ops = {
                                 63                 :                :     /*
                                 64                 :                :      * While io_uring mostly is OK with FDs getting closed while the IO is in
                                 65                 :                :      * flight, that is not true for IOs submitted with IOSQE_ASYNC.
                                 66                 :                :      *
                                 67                 :                :      * See
                                 68                 :                :      * https://postgr.es/m/5ons2rtmwarqqhhexb3dnqulw5rjgwgoct57vpdau4rujlrffj%403fls6d2mkiwc
                                 69                 :                :      */
                                 70                 :                :     .wait_on_fd_before_close = true,
                                 71                 :                : 
                                 72                 :                :     .shmem_size = pgaio_uring_shmem_size,
                                 73                 :                :     .shmem_init = pgaio_uring_shmem_init,
                                 74                 :                :     .init_backend = pgaio_uring_init_backend,
                                 75                 :                : 
                                 76                 :                :     .submit = pgaio_uring_submit,
                                 77                 :                :     .wait_one = pgaio_uring_wait_one,
                                 78                 :                : };
                                 79                 :                : 
                                 80                 :                : /*
                                 81                 :                :  * Per-backend state when using io_method=io_uring
                                 82                 :                :  *
                                 83                 :                :  * Align the whole struct to a cacheline boundary, to prevent false sharing
                                 84                 :                :  * between completion_lock and prior backend's io_uring_ring.
                                 85                 :                :  */
                                 86                 :                : typedef struct pg_attribute_aligned (PG_CACHE_LINE_SIZE)
                                 87                 :                : PgAioUringContext
                                 88                 :                : {
                                 89                 :                :     /*
                                 90                 :                :      * Multiple backends can process completions for this backend's io_uring
                                 91                 :                :      * instance (e.g. when the backend issuing IO is busy doing something
                                 92                 :                :      * else).  To make that safe we have to ensure that only a single backend
                                 93                 :                :      * gets io completions from the io_uring instance at a time.
                                 94                 :                :      */
                                 95                 :                :     LWLock      completion_lock;
                                 96                 :                : 
                                 97                 :                :     struct io_uring io_uring_ring;
                                 98                 :                : } PgAioUringContext;
                                 99                 :                : 
                                100                 :                : /*
                                101                 :                :  * Information about the capabilities that io_uring has.
                                102                 :                :  *
                                103                 :                :  * Depending on liburing and kernel version different features are
                                104                 :                :  * supported. At least for the kernel a kernel version check does not suffice
                                105                 :                :  * as various vendors do backport features to older kernels :(.
                                106                 :                :  */
                                107                 :                : typedef struct PgAioUringCaps
                                108                 :                : {
                                109                 :                :     bool        checked;
                                110                 :                :     /* -1 if io_uring_queue_init_mem() is unsupported */
                                111                 :                :     int         mem_init_size;
                                112                 :                : } PgAioUringCaps;
                                113                 :                : 
                                114                 :                : 
                                115                 :                : /* PgAioUringContexts for all backends */
                                116                 :                : static PgAioUringContext *pgaio_uring_contexts;
                                117                 :                : 
                                118                 :                : /* the current backend's context */
                                119                 :                : static PgAioUringContext *pgaio_my_uring_context;
                                120                 :                : 
                                121                 :                : static PgAioUringCaps pgaio_uring_caps =
                                122                 :                : {
                                123                 :                :     .checked = false,
                                124                 :                :     .mem_init_size = -1,
                                125                 :                : };
                                126                 :                : 
                                127                 :                : static uint32
  172 andres@anarazel.de        128                 :CBC          70 : pgaio_uring_procs(void)
                                129                 :                : {
                                130                 :                :     /*
                                131                 :                :      * We can subtract MAX_IO_WORKERS here as io workers are never used at the
                                132                 :                :      * same time as io_method=io_uring.
                                133                 :                :      */
                                134                 :             70 :     return MaxBackends + NUM_AUXILIARY_PROCS - MAX_IO_WORKERS;
                                135                 :                : }
                                136                 :                : 
                                137                 :                : /*
                                138                 :                :  * Initializes pgaio_uring_caps, unless that's already done.
                                139                 :                :  */
                                140                 :                : static void
   61                           141                 :             12 : pgaio_uring_check_capabilities(void)
                                142                 :                : {
                                143         [ +  + ]:             12 :     if (pgaio_uring_caps.checked)
                                144                 :              7 :         return;
                                145                 :                : 
                                146                 :                :     /*
                                147                 :                :      * By default io_uring creates a shared memory mapping for each io_uring
                                148                 :                :      * instance, leading to a large number of memory mappings. Unfortunately a
                                149                 :                :      * large number of memory mappings slows things down, backend exit is
                                150                 :                :      * particularly affected.  To address that, newer kernels (6.5) support
                                151                 :                :      * using user-provided memory for the memory, by putting the relevant
                                152                 :                :      * memory into shared memory we don't need any additional mappings.
                                153                 :                :      *
                                154                 :                :      * To know whether this is supported, we unfortunately need to probe the
                                155                 :                :      * kernel by trying to create a ring with userspace-provided memory. This
                                156                 :                :      * also has a secondary benefit: We can determine precisely how much
                                157                 :                :      * memory we need for each io_uring instance.
                                158                 :                :      */
                                159                 :                : #if defined(HAVE_LIBURING_QUEUE_INIT_MEM) && defined(IORING_SETUP_NO_MMAP)
                                160                 :                :     {
                                161                 :                :         struct io_uring test_ring;
                                162                 :                :         size_t      ring_size;
                                163                 :                :         void       *ring_ptr;
                                164                 :              5 :         struct io_uring_params p = {0};
                                165                 :                :         int         ret;
                                166                 :                : 
                                167                 :                :         /*
                                168                 :                :          * Liburing does not yet provide an API to query how much memory a
                                169                 :                :          * ring will need. So we over-estimate it here. As the memory is freed
                                170                 :                :          * just below that's small temporary waste of memory.
                                171                 :                :          *
                                172                 :                :          * 1MB is more than enough for rings within io_max_concurrency's
                                173                 :                :          * range.
                                174                 :                :          */
                                175                 :              5 :         ring_size = 1024 * 1024;
                                176                 :                : 
                                177                 :                :         /*
                                178                 :                :          * Hard to believe a system exists where 1MB would not be a multiple
                                179                 :                :          * of the page size. But it's cheap to ensure...
                                180                 :                :          */
                                181                 :              5 :         ring_size -= ring_size % sysconf(_SC_PAGESIZE);
                                182                 :                : 
                                183                 :              5 :         ring_ptr = mmap(NULL, ring_size, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, -1, 0);
                                184         [ -  + ]:              5 :         if (ring_ptr == MAP_FAILED)
   61 andres@anarazel.de        185         [ #  # ]:UBC           0 :             elog(ERROR,
                                186                 :                :                  "mmap(%zu) to determine io_uring_queue_init_mem() support failed: %m",
                                187                 :                :                  ring_size);
                                188                 :                : 
   61 andres@anarazel.de        189                 :CBC           5 :         ret = io_uring_queue_init_mem(io_max_concurrency, &test_ring, &p, ring_ptr, ring_size);
                                190         [ +  - ]:              5 :         if (ret > 0)
                                191                 :                :         {
                                192                 :              5 :             pgaio_uring_caps.mem_init_size = ret;
                                193                 :                : 
                                194         [ +  + ]:              5 :             elog(DEBUG1,
                                195                 :                :                  "can use combined memory mapping for io_uring, each ring needs %d bytes",
                                196                 :                :                  ret);
                                197                 :                : 
                                198                 :                :             /* clean up the created ring, it was just for a test */
                                199                 :              5 :             io_uring_queue_exit(&test_ring);
                                200                 :                :         }
                                201                 :                :         else
                                202                 :                :         {
                                203                 :                :             /*
                                204                 :                :              * There are different reasons for ring creation to fail, but it's
                                205                 :                :              * ok to treat that just as io_uring_queue_init_mem() not being
                                206                 :                :              * supported. We'll report a more detailed error in
                                207                 :                :              * pgaio_uring_shmem_init().
                                208                 :                :              */
   61 andres@anarazel.de        209                 :UBC           0 :             errno = -ret;
                                210         [ #  # ]:              0 :             elog(DEBUG1,
                                211                 :                :                  "cannot use combined memory mapping for io_uring, ring creation failed: %m");
                                212                 :                : 
                                213                 :                :         }
                                214                 :                : 
   61 andres@anarazel.de        215         [ -  + ]:CBC           5 :         if (munmap(ring_ptr, ring_size) != 0)
   61 andres@anarazel.de        216         [ #  # ]:UBC           0 :             elog(ERROR, "munmap() failed: %m");
                                217                 :                :     }
                                218                 :                : #else
                                219                 :                :     {
                                220                 :                :         elog(DEBUG1,
                                221                 :                :              "can't use combined memory mapping for io_uring, kernel or liburing too old");
                                222                 :                :     }
                                223                 :                : #endif
                                224                 :                : 
   61 andres@anarazel.de        225                 :CBC           5 :     pgaio_uring_caps.checked = true;
                                226                 :                : }
                                227                 :                : 
                                228                 :                : /*
                                229                 :                :  * Memory for all PgAioUringContext instances
                                230                 :                :  */
                                231                 :                : static size_t
  172                           232                 :             17 : pgaio_uring_context_shmem_size(void)
                                233                 :                : {
                                234                 :             17 :     return mul_size(pgaio_uring_procs(), sizeof(PgAioUringContext));
                                235                 :                : }
                                236                 :                : 
                                237                 :                : /*
                                238                 :                :  * Memory for the combined memory used by io_uring instances. Returns 0 if
                                239                 :                :  * that is not supported by kernel/liburing.
                                240                 :                :  */
                                241                 :                : static size_t
   61                           242                 :             17 : pgaio_uring_ring_shmem_size(void)
                                243                 :                : {
                                244                 :             17 :     size_t      sz = 0;
                                245                 :                : 
                                246         [ +  - ]:             17 :     if (pgaio_uring_caps.mem_init_size > 0)
                                247                 :                :     {
                                248                 :                :         /*
                                249                 :                :          * Memory for rings needs to be allocated to the page boundary,
                                250                 :                :          * reserve space. Luckily it does not need to be aligned to hugepage
                                251                 :                :          * boundaries, even if huge pages are used.
                                252                 :                :          */
                                253                 :             17 :         sz = add_size(sz, sysconf(_SC_PAGESIZE));
                                254                 :             17 :         sz = add_size(sz, mul_size(pgaio_uring_procs(),
                                255                 :             17 :                                    pgaio_uring_caps.mem_init_size));
                                256                 :                :     }
                                257                 :                : 
                                258                 :             17 :     return sz;
                                259                 :                : }
                                260                 :                : 
                                261                 :                : static size_t
  172                           262                 :             12 : pgaio_uring_shmem_size(void)
                                263                 :                : {
                                264                 :                :     size_t      sz;
                                265                 :                : 
                                266                 :                :     /*
                                267                 :                :      * Kernel and liburing support for various features influences how much
                                268                 :                :      * shmem we need, perform the necessary checks.
                                269                 :                :      */
   61                           270                 :             12 :     pgaio_uring_check_capabilities();
                                271                 :                : 
                                272                 :             12 :     sz = pgaio_uring_context_shmem_size();
                                273                 :             12 :     sz = add_size(sz, pgaio_uring_ring_shmem_size());
                                274                 :                : 
                                275                 :             12 :     return sz;
                                276                 :                : }
                                277                 :                : 
                                278                 :                : static void
  172                           279                 :              5 : pgaio_uring_shmem_init(bool first_time)
                                280                 :                : {
   93 michael@paquier.xyz       281                 :              5 :     int         TotalProcs = pgaio_uring_procs();
                                282                 :                :     bool        found;
                                283                 :                :     char       *shmem;
   61 andres@anarazel.de        284                 :              5 :     size_t      ring_mem_remain = 0;
                                285                 :              5 :     char       *ring_mem_next = 0;
                                286                 :                : 
                                287                 :                :     /*
                                288                 :                :      * We allocate memory for all PgAioUringContext instances and, if
                                289                 :                :      * supported, the memory required for each of the io_uring instances, in
                                290                 :                :      * one ShmemInitStruct().
                                291                 :                :      */
                                292                 :              5 :     shmem = ShmemInitStruct("AioUringContext", pgaio_uring_shmem_size(), &found);
  172                           293         [ -  + ]:              5 :     if (found)
  172 andres@anarazel.de        294                 :UBC           0 :         return;
                                295                 :                : 
   61 andres@anarazel.de        296                 :CBC           5 :     pgaio_uring_contexts = (PgAioUringContext *) shmem;
                                297                 :              5 :     shmem += pgaio_uring_context_shmem_size();
                                298                 :                : 
                                299                 :                :     /* if supported, handle memory alignment / sizing for io_uring memory */
                                300         [ +  - ]:              5 :     if (pgaio_uring_caps.mem_init_size > 0)
                                301                 :                :     {
                                302                 :              5 :         ring_mem_remain = pgaio_uring_ring_shmem_size();
                                303                 :              5 :         ring_mem_next = (char *) shmem;
                                304                 :                : 
                                305                 :                :         /* align to page boundary, see also pgaio_uring_ring_shmem_size() */
                                306                 :              5 :         ring_mem_next = (char *) TYPEALIGN(sysconf(_SC_PAGESIZE), ring_mem_next);
                                307                 :                : 
                                308                 :                :         /* account for alignment */
                                309                 :              5 :         ring_mem_remain -= ring_mem_next - shmem;
                                310                 :              5 :         shmem += ring_mem_next - shmem;
                                311                 :                : 
                                312                 :              5 :         shmem += ring_mem_remain;
                                313                 :                :     }
                                314                 :                : 
  172                           315         [ +  + ]:            705 :     for (int contextno = 0; contextno < TotalProcs; contextno++)
                                316                 :                :     {
                                317                 :            700 :         PgAioUringContext *context = &pgaio_uring_contexts[contextno];
                                318                 :                :         int         ret;
                                319                 :                : 
                                320                 :                :         /*
                                321                 :                :          * Right now a high TotalProcs will cause problems in two ways:
                                322                 :                :          *
                                323                 :                :          * - RLIMIT_NOFILE needs to be big enough to allow all
                                324                 :                :          * io_uring_queue_init() calls to succeed.
                                325                 :                :          *
                                326                 :                :          * - RLIMIT_NOFILE needs to be big enough to still have enough file
                                327                 :                :          * descriptors to satisfy set_max_safe_fds() left over. Or, even
                                328                 :                :          * better, have max_files_per_process left over FDs.
                                329                 :                :          *
                                330                 :                :          * We probably should adjust the soft RLIMIT_NOFILE to ensure that.
                                331                 :                :          *
                                332                 :                :          *
                                333                 :                :          * XXX: Newer versions of io_uring support sharing the workers that
                                334                 :                :          * execute some asynchronous IOs between io_uring instances. It might
                                335                 :                :          * be worth using that - also need to evaluate if that causes
                                336                 :                :          * noticeable additional contention?
                                337                 :                :          */
                                338                 :                : 
                                339                 :                :         /*
                                340                 :                :          * If supported (c.f. pgaio_uring_check_capabilities()), create ring
                                341                 :                :          * with its data in shared memory. Otherwise fall back io_uring
                                342                 :                :          * creating a memory mapping for each ring.
                                343                 :                :          */
                                344                 :                : #if defined(HAVE_LIBURING_QUEUE_INIT_MEM) && defined(IORING_SETUP_NO_MMAP)
   61                           345         [ +  - ]:            700 :         if (pgaio_uring_caps.mem_init_size > 0)
                                346                 :                :         {
                                347                 :            700 :             struct io_uring_params p = {0};
                                348                 :                : 
                                349                 :            700 :             ret = io_uring_queue_init_mem(io_max_concurrency, &context->io_uring_ring, &p, ring_mem_next, ring_mem_remain);
                                350                 :                : 
                                351                 :            700 :             ring_mem_remain -= ret;
                                352                 :            700 :             ring_mem_next += ret;
                                353                 :                :         }
                                354                 :                :         else
                                355                 :                : #endif
                                356                 :                :         {
   61 andres@anarazel.de        357                 :UBC           0 :             ret = io_uring_queue_init(io_max_concurrency, &context->io_uring_ring, 0);
                                358                 :                :         }
                                359                 :                : 
  172 andres@anarazel.de        360         [ -  + ]:CBC         700 :         if (ret < 0)
                                361                 :                :         {
  172 andres@anarazel.de        362                 :UBC           0 :             char       *hint = NULL;
                                363                 :              0 :             int         err = ERRCODE_INTERNAL_ERROR;
                                364                 :                : 
                                365                 :                :             /* add hints for some failures that errno explains sufficiently */
                                366         [ #  # ]:              0 :             if (-ret == EPERM)
                                367                 :                :             {
                                368                 :              0 :                 err = ERRCODE_INSUFFICIENT_PRIVILEGE;
                                369                 :              0 :                 hint = _("Check if io_uring is disabled via /proc/sys/kernel/io_uring_disabled.");
                                370                 :                :             }
                                371         [ #  # ]:              0 :             else if (-ret == EMFILE)
                                372                 :                :             {
                                373                 :              0 :                 err = ERRCODE_INSUFFICIENT_RESOURCES;
                                374                 :              0 :                 hint = psprintf(_("Consider increasing \"ulimit -n\" to at least %d."),
                                375                 :                :                                 TotalProcs + max_files_per_process);
                                376                 :                :             }
                                377         [ #  # ]:              0 :             else if (-ret == ENOSYS)
                                378                 :                :             {
                                379                 :              0 :                 err = ERRCODE_FEATURE_NOT_SUPPORTED;
   11 peter@eisentraut.org      380                 :              0 :                 hint = _("The kernel does not support io_uring.");
                                381                 :                :             }
                                382                 :                : 
                                383                 :                :             /* update errno to allow %m to work */
  172 andres@anarazel.de        384                 :              0 :             errno = -ret;
                                385                 :                : 
                                386   [ #  #  #  # ]:              0 :             ereport(ERROR,
                                387                 :                :                     errcode(err),
                                388                 :                :                     errmsg("could not setup io_uring queue: %m"),
                                389                 :                :                     hint != NULL ? errhint("%s", hint) : 0);
                                390                 :                :         }
                                391                 :                : 
  172 andres@anarazel.de        392                 :CBC         700 :         LWLockInitialize(&context->completion_lock, LWTRANCHE_AIO_URING_COMPLETION);
                                393                 :                :     }
                                394                 :                : }
                                395                 :                : 
                                396                 :                : static void
                                397                 :             31 : pgaio_uring_init_backend(void)
                                398                 :                : {
                                399         [ -  + ]:             31 :     Assert(MyProcNumber < pgaio_uring_procs());
                                400                 :                : 
                                401                 :             31 :     pgaio_my_uring_context = &pgaio_uring_contexts[MyProcNumber];
                                402                 :             31 : }
                                403                 :                : 
                                404                 :                : static int
                                405                 :            391 : pgaio_uring_submit(uint16 num_staged_ios, PgAioHandle **staged_ios)
                                406                 :                : {
                                407                 :            391 :     struct io_uring *uring_instance = &pgaio_my_uring_context->io_uring_ring;
                                408                 :            391 :     int         in_flight_before = dclist_count(&pgaio_my_backend->in_flight_ios);
                                409                 :                : 
                                410         [ -  + ]:            391 :     Assert(num_staged_ios <= PGAIO_SUBMIT_BATCH_SIZE);
                                411                 :                : 
                                412         [ +  + ]:            783 :     for (int i = 0; i < num_staged_ios; i++)
                                413                 :                :     {
                                414                 :            392 :         PgAioHandle *ioh = staged_ios[i];
                                415                 :                :         struct io_uring_sqe *sqe;
                                416                 :                : 
                                417                 :            392 :         sqe = io_uring_get_sqe(uring_instance);
                                418                 :                : 
                                419         [ -  + ]:            392 :         if (!sqe)
  172 andres@anarazel.de        420         [ #  # ]:UBC           0 :             elog(ERROR, "io_uring submission queue is unexpectedly full");
                                421                 :                : 
  172 andres@anarazel.de        422                 :CBC         392 :         pgaio_io_prepare_submit(ioh);
                                423                 :            392 :         pgaio_uring_sq_from_io(ioh, sqe);
                                424                 :                : 
                                425                 :                :         /*
                                426                 :                :          * io_uring executes IO in process context if possible. That's
                                427                 :                :          * generally good, as it reduces context switching. When performing a
                                428                 :                :          * lot of buffered IO that means that copying between page cache and
                                429                 :                :          * userspace memory happens in the foreground, as it can't be
                                430                 :                :          * offloaded to DMA hardware as is possible when using direct IO. When
                                431                 :                :          * executing a lot of buffered IO this causes io_uring to be slower
                                432                 :                :          * than worker mode, as worker mode parallelizes the copying. io_uring
                                433                 :                :          * can be told to offload work to worker threads instead.
                                434                 :                :          *
                                435                 :                :          * If an IO is buffered IO and we already have IOs in flight or
                                436                 :                :          * multiple IOs are being submitted, we thus tell io_uring to execute
                                437                 :                :          * the IO in the background. We don't do so for the first few IOs
                                438                 :                :          * being submitted as executing in this process' context has lower
                                439                 :                :          * latency.
                                440                 :                :          */
                                441   [ -  +  -  - ]:            392 :         if (in_flight_before > 4 && (ioh->flags & PGAIO_HF_BUFFERED))
  172 andres@anarazel.de        442                 :UBC           0 :             io_uring_sqe_set_flags(sqe, IOSQE_ASYNC);
                                443                 :                : 
  172 andres@anarazel.de        444                 :CBC         392 :         in_flight_before++;
                                445                 :                :     }
                                446                 :                : 
                                447                 :                :     while (true)
  172 andres@anarazel.de        448                 :UBC           0 :     {
                                449                 :                :         int         ret;
                                450                 :                : 
  172 andres@anarazel.de        451                 :CBC         391 :         pgstat_report_wait_start(WAIT_EVENT_AIO_IO_URING_SUBMIT);
                                452                 :            391 :         ret = io_uring_submit(uring_instance);
                                453                 :            391 :         pgstat_report_wait_end();
                                454                 :                : 
                                455         [ -  + ]:            391 :         if (ret == -EINTR)
                                456                 :                :         {
  172 andres@anarazel.de        457         [ #  # ]:UBC           0 :             pgaio_debug(DEBUG3,
                                458                 :                :                         "aio method uring: submit EINTR, nios: %d",
                                459                 :                :                         num_staged_ios);
                                460                 :                :         }
  172 andres@anarazel.de        461         [ -  + ]:CBC         391 :         else if (ret < 0)
                                462                 :                :         {
                                463                 :                :             /*
                                464                 :                :              * The io_uring_enter() manpage suggests that the appropriate
                                465                 :                :              * reaction to EAGAIN is:
                                466                 :                :              *
                                467                 :                :              * "The application should wait for some completions and try
                                468                 :                :              * again"
                                469                 :                :              *
                                470                 :                :              * However, it seems unlikely that that would help in our case, as
                                471                 :                :              * we apply a low limit to the number of outstanding IOs and thus
                                472                 :                :              * also outstanding completions, making it unlikely that we'd get
                                473                 :                :              * EAGAIN while the OS is in good working order.
                                474                 :                :              *
                                475                 :                :              * Additionally, it would be problematic to just wait here, our
                                476                 :                :              * caller might hold critical locks. It'd possibly lead to
                                477                 :                :              * delaying the crash-restart that seems likely to occur when the
                                478                 :                :              * kernel is under such heavy memory pressure.
                                479                 :                :              *
                                480                 :                :              * Update errno to allow %m to work.
                                481                 :                :              */
  172 andres@anarazel.de        482                 :UBC           0 :             errno = -ret;
                                483         [ #  # ]:              0 :             elog(PANIC, "io_uring submit failed: %m");
                                484                 :                :         }
  172 andres@anarazel.de        485         [ -  + ]:CBC         391 :         else if (ret != num_staged_ios)
                                486                 :                :         {
                                487                 :                :             /* likely unreachable, but if it is, we would need to re-submit */
  172 andres@anarazel.de        488         [ #  # ]:UBC           0 :             elog(PANIC, "io_uring submit submitted only %d of %d",
                                489                 :                :                  ret, num_staged_ios);
                                490                 :                :         }
                                491                 :                :         else
                                492                 :                :         {
  172 andres@anarazel.de        493         [ -  + ]:CBC         391 :             pgaio_debug(DEBUG4,
                                494                 :                :                         "aio method uring: submitted %d IOs",
                                495                 :                :                         num_staged_ios);
                                496                 :            391 :             break;
                                497                 :                :         }
                                498                 :                :     }
                                499                 :                : 
                                500                 :            391 :     return num_staged_ios;
                                501                 :                : }
                                502                 :                : 
                                503                 :                : static void
  158 melanieplageman@gmai      504                 :           1112 : pgaio_uring_completion_error_callback(void *arg)
                                505                 :                : {
                                506                 :                :     ProcNumber  owner;
                                507                 :                :     PGPROC     *owner_proc;
                                508                 :                :     int32       owner_pid;
                                509                 :           1112 :     PgAioHandle *ioh = arg;
                                510                 :                : 
                                511         [ +  + ]:           1112 :     if (!ioh)
                                512                 :            334 :         return;
                                513                 :                : 
                                514                 :                :     /* No need for context if a backend is completing the IO for itself */
                                515         [ +  + ]:            778 :     if (ioh->owner_procno == MyProcNumber)
                                516                 :            771 :         return;
                                517                 :                : 
                                518                 :              7 :     owner = ioh->owner_procno;
                                519                 :              7 :     owner_proc = GetPGProcByNumber(owner);
                                520                 :              7 :     owner_pid = owner_proc->pid;
                                521                 :                : 
                                522                 :              7 :     errcontext("completing I/O on behalf of process %d", owner_pid);
                                523                 :                : }
                                524                 :                : 
                                525                 :                : static void
  172 andres@anarazel.de        526                 :            383 : pgaio_uring_drain_locked(PgAioUringContext *context)
                                527                 :                : {
                                528                 :                :     int         ready;
                                529                 :                :     int         orig_ready;
  158 melanieplageman@gmai      530                 :            383 :     ErrorContextCallback errcallback = {0};
                                531                 :                : 
  172 andres@anarazel.de        532         [ -  + ]:            383 :     Assert(LWLockHeldByMeInMode(&context->completion_lock, LW_EXCLUSIVE));
                                533                 :                : 
  158 melanieplageman@gmai      534                 :            383 :     errcallback.callback = pgaio_uring_completion_error_callback;
                                535                 :            383 :     errcallback.previous = error_context_stack;
                                536                 :            383 :     error_context_stack = &errcallback;
                                537                 :                : 
                                538                 :                :     /*
                                539                 :                :      * Don't drain more events than available right now. Otherwise it's
                                540                 :                :      * plausible that one backend could get stuck, for a while, receiving CQEs
                                541                 :                :      * without actually processing them.
                                542                 :                :      */
  172 andres@anarazel.de        543                 :            383 :     orig_ready = ready = io_uring_cq_ready(&context->io_uring_ring);
                                544                 :                : 
                                545         [ +  + ]:            766 :     while (ready > 0)
                                546                 :                :     {
                                547                 :                :         struct io_uring_cqe *cqes[PGAIO_MAX_LOCAL_COMPLETED_IO];
                                548                 :                :         uint32      ncqes;
                                549                 :                : 
                                550                 :            383 :         START_CRIT_SECTION();
                                551                 :                :         ncqes =
                                552                 :            383 :             io_uring_peek_batch_cqe(&context->io_uring_ring,
                                553                 :                :                                     cqes,
                                554                 :            383 :                                     Min(PGAIO_MAX_LOCAL_COMPLETED_IO, ready));
                                555         [ -  + ]:            383 :         Assert(ncqes <= ready);
                                556                 :                : 
                                557                 :            383 :         ready -= ncqes;
                                558                 :                : 
                                559         [ +  + ]:            775 :         for (int i = 0; i < ncqes; i++)
                                560                 :                :         {
                                561                 :            392 :             struct io_uring_cqe *cqe = cqes[i];
                                562                 :                :             PgAioHandle *ioh;
                                563                 :                : 
                                564                 :            392 :             ioh = io_uring_cqe_get_data(cqe);
  158 melanieplageman@gmai      565                 :            392 :             errcallback.arg = ioh;
  172 andres@anarazel.de        566                 :            392 :             io_uring_cqe_seen(&context->io_uring_ring, cqe);
                                567                 :                : 
                                568                 :            392 :             pgaio_io_process_completion(ioh, cqe->res);
  158 melanieplageman@gmai      569                 :            392 :             errcallback.arg = NULL;
                                570                 :                :         }
                                571                 :                : 
  172 andres@anarazel.de        572         [ -  + ]:            383 :         END_CRIT_SECTION();
                                573                 :                : 
                                574         [ +  + ]:            383 :         pgaio_debug(DEBUG3,
                                575                 :                :                     "drained %d/%d, now expecting %d",
                                576                 :                :                     ncqes, orig_ready, io_uring_cq_ready(&context->io_uring_ring));
                                577                 :                :     }
                                578                 :                : 
  158 melanieplageman@gmai      579                 :            383 :     error_context_stack = errcallback.previous;
  172 andres@anarazel.de        580                 :            383 : }
                                581                 :                : 
                                582                 :                : static void
                                583                 :            383 : pgaio_uring_wait_one(PgAioHandle *ioh, uint64 ref_generation)
                                584                 :                : {
                                585                 :                :     PgAioHandleState state;
                                586                 :            383 :     ProcNumber  owner_procno = ioh->owner_procno;
                                587                 :            383 :     PgAioUringContext *owner_context = &pgaio_uring_contexts[owner_procno];
                                588                 :                :     bool        expect_cqe;
                                589                 :            383 :     int         waited = 0;
                                590                 :                : 
                                591                 :                :     /*
                                592                 :                :      * XXX: It would be nice to have a smarter locking scheme, nearly all the
                                593                 :                :      * time the backend owning the ring will consume the completions, making
                                594                 :                :      * the locking unnecessarily expensive.
                                595                 :                :      */
                                596                 :            383 :     LWLockAcquire(&owner_context->completion_lock, LW_EXCLUSIVE);
                                597                 :                : 
                                598                 :                :     while (true)
                                599                 :                :     {
                                600         [ +  + ]:           1149 :         pgaio_debug_io(DEBUG3, ioh,
                                601                 :                :                        "wait_one io_gen: %" PRIu64 ", ref_gen: %" PRIu64 ", cycle %d",
                                602                 :                :                        ioh->generation,
                                603                 :                :                        ref_generation,
                                604                 :                :                        waited);
                                605                 :                : 
                                606         [ +  + ]:            766 :         if (pgaio_io_was_recycled(ioh, ref_generation, &state) ||
                                607         [ +  + ]:            386 :             state != PGAIO_HS_SUBMITTED)
                                608                 :                :         {
                                609                 :                :             /* the IO was completed by another backend */
                                610                 :                :             break;
                                611                 :                :         }
                                612         [ +  - ]:            383 :         else if (io_uring_cq_ready(&owner_context->io_uring_ring))
                                613                 :                :         {
                                614                 :                :             /* no need to wait in the kernel, io_uring has a completion */
                                615                 :            383 :             expect_cqe = true;
                                616                 :                :         }
                                617                 :                :         else
                                618                 :                :         {
                                619                 :                :             int         ret;
                                620                 :                :             struct io_uring_cqe *cqes;
                                621                 :                : 
                                622                 :                :             /* need to wait in the kernel */
  172 andres@anarazel.de        623                 :UBC           0 :             pgstat_report_wait_start(WAIT_EVENT_AIO_IO_URING_EXECUTION);
                                624                 :              0 :             ret = io_uring_wait_cqes(&owner_context->io_uring_ring, &cqes, 1, NULL, NULL);
                                625                 :              0 :             pgstat_report_wait_end();
                                626                 :                : 
                                627         [ #  # ]:              0 :             if (ret == -EINTR)
                                628                 :                :             {
                                629                 :              0 :                 continue;
                                630                 :                :             }
                                631         [ #  # ]:              0 :             else if (ret != 0)
                                632                 :                :             {
                                633                 :                :                 /* see comment after io_uring_submit() */
                                634                 :              0 :                 errno = -ret;
                                635         [ #  # ]:              0 :                 elog(PANIC, "io_uring wait failed: %m");
                                636                 :                :             }
                                637                 :                :             else
                                638                 :                :             {
                                639         [ #  # ]:              0 :                 Assert(cqes != NULL);
                                640                 :              0 :                 expect_cqe = true;
                                641                 :              0 :                 waited++;
                                642                 :                :             }
                                643                 :                :         }
                                644                 :                : 
  172 andres@anarazel.de        645         [ +  - ]:CBC         383 :         if (expect_cqe)
                                646                 :                :         {
                                647                 :            383 :             pgaio_uring_drain_locked(owner_context);
                                648                 :                :         }
                                649                 :                :     }
                                650                 :                : 
                                651                 :            383 :     LWLockRelease(&owner_context->completion_lock);
                                652                 :                : 
                                653         [ +  + ]:            383 :     pgaio_debug(DEBUG3,
                                654                 :                :                 "wait_one with %d sleeps",
                                655                 :                :                 waited);
                                656                 :            383 : }
                                657                 :                : 
                                658                 :                : static void
                                659                 :            392 : pgaio_uring_sq_from_io(PgAioHandle *ioh, struct io_uring_sqe *sqe)
                                660                 :                : {
                                661                 :                :     struct iovec *iov;
                                662                 :                : 
   10                           663   [ +  -  -  - ]:            392 :     switch ((PgAioOp) ioh->op)
                                664                 :                :     {
  172                           665                 :            392 :         case PGAIO_OP_READV:
                                666                 :            392 :             iov = &pgaio_ctl->iovecs[ioh->iovec_off];
                                667         [ +  + ]:            392 :             if (ioh->op_data.read.iov_length == 1)
                                668                 :                :             {
                                669                 :            382 :                 io_uring_prep_read(sqe,
                                670                 :                :                                    ioh->op_data.read.fd,
                                671                 :                :                                    iov->iov_base,
                                672                 :            382 :                                    iov->iov_len,
                                673                 :            382 :                                    ioh->op_data.read.offset);
                                674                 :                :             }
                                675                 :                :             else
                                676                 :                :             {
                                677                 :             10 :                 io_uring_prep_readv(sqe,
                                678                 :                :                                     ioh->op_data.read.fd,
                                679                 :                :                                     iov,
                                680                 :             10 :                                     ioh->op_data.read.iov_length,
                                681                 :             10 :                                     ioh->op_data.read.offset);
                                682                 :                : 
                                683                 :                :             }
                                684                 :            392 :             break;
                                685                 :                : 
  172 andres@anarazel.de        686                 :UBC           0 :         case PGAIO_OP_WRITEV:
                                687                 :              0 :             iov = &pgaio_ctl->iovecs[ioh->iovec_off];
                                688         [ #  # ]:              0 :             if (ioh->op_data.write.iov_length == 1)
                                689                 :                :             {
                                690                 :              0 :                 io_uring_prep_write(sqe,
                                691                 :                :                                     ioh->op_data.write.fd,
                                692                 :              0 :                                     iov->iov_base,
                                693                 :              0 :                                     iov->iov_len,
                                694                 :              0 :                                     ioh->op_data.write.offset);
                                695                 :                :             }
                                696                 :                :             else
                                697                 :                :             {
                                698                 :              0 :                 io_uring_prep_writev(sqe,
                                699                 :                :                                      ioh->op_data.write.fd,
                                700                 :                :                                      iov,
                                701                 :              0 :                                      ioh->op_data.write.iov_length,
                                702                 :              0 :                                      ioh->op_data.write.offset);
                                703                 :                :             }
                                704                 :              0 :             break;
                                705                 :                : 
                                706                 :              0 :         case PGAIO_OP_INVALID:
                                707         [ #  # ]:              0 :             elog(ERROR, "trying to prepare invalid IO operation for execution");
                                708                 :                :     }
                                709                 :                : 
  172 andres@anarazel.de        710                 :CBC         392 :     io_uring_sqe_set_data(sqe, ioh);
                                711                 :            392 : }
                                712                 :                : 
                                713                 :                : #endif                          /* IOMETHOD_IO_URING_ENABLED */
        

Generated by: LCOV version 2.4-beta