From patchwork Thu Dec 28 20:35:41 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Steven Rostedt X-Patchwork-Id: 13506210 Received: from smtp.kernel.org (aws-us-west-2-korg-mail-1.web.codeaurora.org [10.30.226.201]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.subspace.kernel.org (Postfix) with ESMTPS id D6408107BE for ; Thu, 28 Dec 2023 20:36:26 +0000 (UTC) Received: by smtp.kernel.org (Postfix) with ESMTPSA id 42263C433BA; Thu, 28 Dec 2023 20:36:26 +0000 (UTC) Received: from rostedt by gandalf with local (Exim 4.97) (envelope-from ) id 1rIx80-00000000Dtk-04ze; Thu, 28 Dec 2023 15:37:16 -0500 From: Steven Rostedt To: linux-trace-devel@vger.kernel.org Cc: "Steven Rostedt (Google)" Subject: [PATCH 19/23] libtracefs: Add ring buffer memory mapping APIs Date: Thu, 28 Dec 2023 15:35:41 -0500 Message-ID: <20231228203714.53294-20-rostedt@goodmis.org> X-Mailer: git-send-email 2.42.0 In-Reply-To: <20231228203714.53294-1-rostedt@goodmis.org> References: <20231228203714.53294-1-rostedt@goodmis.org> Precedence: bulk X-Mailing-List: linux-trace-devel@vger.kernel.org List-Id: List-Subscribe: List-Unsubscribe: MIME-Version: 1.0 From: "Steven Rostedt (Google)" Add the following APIs: tracefs_cpu_open_mapped() tracefs_cpu_is_mapped() tracefs_cpu_map() tracefs_cpu_unmap() This will allow applications to choose to memory map the tracing ring buffer if it is supported. This will improve the performance of tracefs_cpu_read() and tracefs_cpu_read_buf(), but it is not done by default because it will also hurt the performance of tracefs_cpu_buffered_read() and tracefs_cpu_buffered_read_buf() as those use splicing, and with the ring buffer memory mapped, the splice has to do a copy instead of a copyless subbuffer move. Signed-off-by: Steven Rostedt (Google) --- Documentation/libtracefs-cpu-map.txt | 194 +++++++++++++++++++++++++++ Documentation/libtracefs.txt | 7 + include/tracefs-local.h | 6 + include/tracefs.h | 7 + samples/Makefile | 1 + src/Makefile | 1 + src/tracefs-mmap.c | 190 ++++++++++++++++++++++++++ src/tracefs-record.c | 61 +++++++++ 8 files changed, 467 insertions(+) create mode 100644 Documentation/libtracefs-cpu-map.txt create mode 100644 src/tracefs-mmap.c diff --git a/Documentation/libtracefs-cpu-map.txt b/Documentation/libtracefs-cpu-map.txt new file mode 100644 index 000000000000..9bcd29715795 --- /dev/null +++ b/Documentation/libtracefs-cpu-map.txt @@ -0,0 +1,194 @@ +libtracefs(3) +============= + +NAME +---- +tracefs_cpu_open_mapped, tracefs_cpu_is_mapped, tracefs_cpu_map, tracefs_cpu_unmap - Memory mapping of the ring buffer + +SYNOPSIS +-------- +[verse] +-- +*#include * + +bool *tracefs_cpu_is_mapped*(struct tracefs_cpu pass:[*]tcpu); +int *tracefs_cpu_map*(struct tracefs_cpu pass:[*]tcpu); +void *tracefs_cpu_unmap*(struct tracefs_cpu pass:[*]tcpu); +struct tracefs_cpu pass:[*]*tracefs_cpu_open_mapped*(struct tracefs_instance pass:[*]instance, + int cpu, bool nonblock); +-- + +DESCRIPTION +----------- +If the trace_pipe_raw supports memory mapping, this is usually a more efficient +method to stream data from the kernel ring buffer than by reading it, as it does +not require copying the memory that is being read. + +If memory mapping is supported by the kernel and the application asks to use the +memory mapping via either *tracefs_cpu_map()* or by *tracefs_cpu_open_mapped()* +then the functions *tracefs_cpu_read*(3) and *tracefs_cpu_read_buf*(3) will use +the mapping directly instead of calling the read system call. + +Note, mapping can also slow down *tracefs_cpu_buffered_read*(3) and +*tracefs_cpu_buffered_read_buf*(3), as those use splice piping and when the +kernel ring buffer is memory mapped, splice does a copy instead of using the +ring buffer directly. Thus care must be used when determining to map the +ring buffer or not, and why it does not get mapped by default. + +The *tracefs_cpu_is_mapped()* function will return true if _tcpu_ currently has +its ring buffer memory mapped and false otherwise. This does not return whether or +not that the kernel supports memory mapping, but that can usually be determined +by calling *tracefs_cpu_map()*. + +The *tracefs_cpu_map()* function will attempt to map the ring buffer associated +to _tcpu_ if it is not already mapped. + +The *tracefs_cpu_unmap()* function will unmap the ring buffer associated to +_tcpu_ if it is mapped. + +The *tracefs_cpu_open_mapped()* is equivalent to calling *tracefs_cpu_open*(3) followed +by *tracefs_cpu_map()* on the returned _tcpu_ of *tracefs_cpu_open*(3). Note, this +will still succeed if the mapping fails, in which case it acts the same as +*tracefs_cpu_open*(3). If knowing if the mapping succeed or not, *tracefs_cpu_is_mapped()* +should be called on the return _tcpu_. + +RETURN VALUE +------------ +*tracefs_cpu_is_mapped()* returns true if the given _tcpu_ has its ring buffer +memory mapped or false otherwise. + +*tracefs_cpu_map()* returns 0 on success and -1 on error in mapping. If 0 is +returned then *tracefs_cpu_is_mapped()* will return true afterward, or false +if the mapping failed. + +*tracefs_cpu_open_mapped()* returns an allocated tracefs_cpu on success of creation +regardless if it succeed in mapping the ring buffer or not. It returns NULL for +the same reasons *tracefs_cpu_open*(3) returns NULL. If success of mapping is +to be known, then calling *tracefs_cpu_is_mapped()* afterward is required. + +EXAMPLE +------- +[source,c] +-- +#include +#include +#include + +static void read_page(struct tep_handle *tep, struct kbuffer *kbuf) +{ + static struct trace_seq seq; + struct tep_record record; + + if (seq.buffer) + trace_seq_reset(&seq); + else + trace_seq_init(&seq); + + while ((record.data = kbuffer_read_event(kbuf, &record.ts))) { + record.size = kbuffer_event_size(kbuf); + kbuffer_next_event(kbuf, NULL); + tep_print_event(tep, &seq, &record, + "%s-%d %9d\t%s: %s\n", + TEP_PRINT_COMM, + TEP_PRINT_PID, + TEP_PRINT_TIME, + TEP_PRINT_NAME, + TEP_PRINT_INFO); + trace_seq_do_printf(&seq); + trace_seq_reset(&seq); + } +} + +int main (int argc, char **argv) +{ + struct tracefs_cpu *tcpu; + struct tep_handle *tep; + struct kbuffer *kbuf; + bool mapped; + int cpu; + + if (argc < 2 || !isdigit(argv[1][0])) { + printf("usage: %s cpu\n\n", argv[0]); + exit(-1); + } + + cpu = atoi(argv[1]); + + tep = tracefs_local_events(NULL); + if (!tep) { + perror("Reading trace event formats"); + exit(-1); + } + + tcpu = tracefs_cpu_open_mapped(NULL, cpu, 0); + if (!tcpu) { + perror("Open CPU 0 file"); + exit(-1); + } + + /* + * If this kernel supports mapping, use normal read, + * otherwise use the piped buffer read. + */ + mapped = tracefs_cpu_is_mapped(tcpu); + if (!mapped) + printf("Was not able to map, falling back to buffered read\n"); + while ((kbuf = mapped ? tracefs_cpu_read_buf(tcpu, true) : + tracefs_cpu_buffered_read_buf(tcpu, true))) { + read_page(tep, kbuf); + } + + kbuf = tracefs_cpu_flush_buf(tcpu); + if (kbuf) + read_page(tep, kbuf); + + tracefs_cpu_close(tcpu); + tep_free(tep); + + return 0; +} +-- + +FILES +----- +[verse] +-- +*tracefs.h* + Header file to include in order to have access to the library APIs. +*-ltracefs* + Linker switch to add when building a program that uses the library. +-- + +SEE ALSO +-------- +*tracefs_cpu_open*(3), +*tracefs_cpu_read*(3), +*tracefs_cpu_read_buf*(3), +*tracefs_cpu_buffered_read*(3), +*tracefs_cpu_buffered_read_buf*(3), +*libtracefs*(3), +*libtraceevent*(3), +*trace-cmd*(1) + +AUTHOR +------ +[verse] +-- +*Steven Rostedt* +-- +REPORTING BUGS +-------------- +Report bugs to + +LICENSE +------- +libtracefs is Free Software licensed under the GNU LGPL 2.1 + +RESOURCES +--------- +https://git.kernel.org/pub/scm/libs/libtrace/libtracefs.git/ + +COPYING +------- +Copyright \(C) 2022 Google, Inc. Free use of this software is granted under +the terms of the GNU Public License (GPL). diff --git a/Documentation/libtracefs.txt b/Documentation/libtracefs.txt index 1962b3b34622..2f2b4d488942 100644 --- a/Documentation/libtracefs.txt +++ b/Documentation/libtracefs.txt @@ -124,6 +124,13 @@ Trace stream: ssize_t *tracefs_trace_pipe_print*(struct tracefs_instance pass:[*]_instance_, int _flags_); void *tracefs_trace_pipe_stop*(struct tracefs_instance pass:[*]_instance_); +Memory mapping the ring buffer: + bool *tracefs_cpu_is_mapped*(struct tracefs_cpu pass:[*]tcpu); + int *tracefs_cpu_map*(struct tracefs_cpu pass:[*]tcpu); + void *tracefs_cpu_unmap*(struct tracefs_cpu pass:[*]tcpu); + struct tracefs_cpu pass:[*]*tracefs_cpu_open_mapped*(struct tracefs_instance pass:[*]instance, + int cpu, bool nonblock); + Trace options: const struct tracefs_options_mask pass:[*]*tracefs_options_get_supported*(struct tracefs_instance pass:[*]_instance_); bool *tracefs_option_is_supported*(struct tracefs_instance pass:[*]_instance_, enum tracefs_option_id _id_); diff --git a/include/tracefs-local.h b/include/tracefs-local.h index 9cae73c8b806..ffc9d33b1796 100644 --- a/include/tracefs-local.h +++ b/include/tracefs-local.h @@ -6,6 +6,7 @@ #ifndef _TRACE_FS_LOCAL_H #define _TRACE_FS_LOCAL_H +#include #include #define __hidden __attribute__((visibility ("hidden"))) @@ -116,6 +117,11 @@ int trace_append_filter(char **filter, unsigned int *state, enum tracefs_compare compare, const char *val); +void *trace_mmap(int fd, struct kbuffer *kbuf); +void trace_unmap(void *mapping); +int trace_mmap_load_subbuf(void *mapping, struct kbuffer *kbuf); +int trace_mmap_read(void *mapping, void *buffer); + struct tracefs_synth *synth_init_from(struct tep_handle *tep, const char *start_system, const char *start_event); diff --git a/include/tracefs.h b/include/tracefs.h index c66bfd2edf0b..d20d733b7d92 100644 --- a/include/tracefs.h +++ b/include/tracefs.h @@ -673,6 +673,13 @@ struct kbuffer *tracefs_cpu_flush_buf(struct tracefs_cpu *tcpu); int tracefs_cpu_flush_write(struct tracefs_cpu *tcpu, int wfd); int tracefs_cpu_pipe(struct tracefs_cpu *tcpu, int wfd, bool nonblock); +/* Memory mapping of ring buffer */ +bool tracefs_cpu_is_mapped(struct tracefs_cpu *tcpu); +int tracefs_cpu_map(struct tracefs_cpu *tcpu); +void tracefs_cpu_unmap(struct tracefs_cpu *tcpu); +struct tracefs_cpu *tracefs_cpu_open_mapped(struct tracefs_instance *instance, + int cpu, bool nonblock); + /* Mapping vsocket cids to pids using tracing */ int tracefs_instance_find_cid_pid(struct tracefs_instance *instance, int cid); int tracefs_find_cid_pid(int cid); diff --git a/samples/Makefile b/samples/Makefile index 77739c8b0aa7..81c8006f823e 100644 --- a/samples/Makefile +++ b/samples/Makefile @@ -26,6 +26,7 @@ EXAMPLES += guest EXAMPLES += cpu-buf EXAMPLES += instances-stat EXAMPLES += instances-subbuf +EXAMPLES += cpu-map TARGETS := TARGETS += sqlhist diff --git a/src/Makefile b/src/Makefile index faa3b25c4002..be81059ce10a 100644 --- a/src/Makefile +++ b/src/Makefile @@ -16,6 +16,7 @@ OBJS += tracefs-dynevents.o OBJS += tracefs-eprobes.o OBJS += tracefs-uprobes.o OBJS += tracefs-record.o +OBJS += tracefs-mmap.o ifeq ($(VSOCK_DEFINED), 1) OBJS += tracefs-vsock.o endif diff --git a/src/tracefs-mmap.c b/src/tracefs-mmap.c new file mode 100644 index 000000000000..5675028acc51 --- /dev/null +++ b/src/tracefs-mmap.c @@ -0,0 +1,190 @@ +// SPDX-License-Identifier: LGPL-2.1 +/* + * Copyright (C) 2023 Google Inc, Steven Rostedt + */ +#include +#include +#include +#include +#include +#include "tracefs-local.h" + +struct trace_buffer_meta { + unsigned long entries; + unsigned long overrun; + unsigned long read; + + unsigned long subbufs_touched; + unsigned long subbufs_lost; + unsigned long subbufs_read; + + struct { + unsigned long lost_events; /* Events lost at the time of the reader swap */ + __u32 id; /* Reader subbuf ID from 0 to nr_subbufs - 1 */ + __u32 read; /* Number of bytes read on the reader subbuf */ + } reader; + + __u32 subbuf_size; /* Size of each subbuf including the header */ + __u32 nr_subbufs; /* Number of subbufs in the ring-buffer */ + + __u32 meta_page_size; /* Size of the meta-page */ + __u32 meta_struct_len; /* Len of this struct */ +}; + +#define TRACE_MMAP_IOCTL_GET_READER _IO('T', 0x1) + +struct trace_mmap { + struct trace_buffer_meta *map; + struct kbuffer *kbuf; + void *data; + int *data_pages; + int fd; + int last_idx; + int last_read; + int meta_len; + int data_len; +}; + +/** + * trace_mmap - try to mmap the ring buffer + * @fd: The file descriptor to the trace_pipe_raw file + * + * Will try to mmap the ring buffer if it is supported, and + * if not, will return NULL, otherwise it returns a descriptor + * to handle the mapping. + */ +void *trace_mmap(int fd, struct kbuffer *kbuf) +{ + struct trace_mmap *tmap; + int page_size; + void *meta; + void *data; + + page_size = getpagesize(); + meta = mmap(NULL, page_size, PROT_READ, MAP_SHARED, fd, 0); + if (meta == MAP_FAILED) + return NULL; + + tmap = calloc(1, sizeof(*tmap)); + if (!tmap) { + munmap(meta, page_size); + return NULL; + } + + tmap->kbuf = kbuffer_dup(kbuf); + if (!tmap->kbuf) { + munmap(meta, page_size); + free(tmap); + } + + tmap->fd = fd; + + tmap->map = meta; + tmap->meta_len = tmap->map->meta_page_size; + + if (tmap->meta_len > page_size) { + munmap(meta, page_size); + meta = mmap(NULL, tmap->meta_len, PROT_READ, MAP_SHARED, fd, 0); + if (meta == MAP_FAILED) { + kbuffer_free(tmap->kbuf); + free(tmap); + return NULL; + } + tmap->map = meta; + } + + tmap->data_pages = meta + tmap->meta_len; + + tmap->data_len = tmap->map->subbuf_size * tmap->map->nr_subbufs; + + tmap->data = mmap(NULL, tmap->data_len, PROT_READ, MAP_SHARED, + fd, tmap->meta_len); + if (tmap->data == MAP_FAILED) { + munmap(meta, tmap->meta_len); + kbuffer_free(tmap->kbuf); + free(tmap); + return NULL; + } + + tmap->last_idx = tmap->map->reader.id; + + data = tmap->data + tmap->map->subbuf_size * tmap->last_idx; + kbuffer_load_subbuffer(kbuf, data); + + return tmap; +} + +void trace_unmap(void *mapping) +{ + struct trace_mmap *tmap = mapping; + + munmap(tmap->data, tmap->data_len); + munmap(tmap->map, tmap->meta_len); + kbuffer_free(tmap->kbuf); + free(tmap); +} + +int trace_mmap_load_subbuf(void *mapping, struct kbuffer *kbuf) +{ + struct trace_mmap *tmap = mapping; + void *data; + int id; + + id = tmap->map->reader.id; + data = tmap->data + tmap->map->subbuf_size * id; + + /* + * If kbuf doesn't point to the current sub-buffer + * just load it and return. + */ + if (data != kbuffer_subbuffer(kbuf)) { + kbuffer_load_subbuffer(kbuf, data); + return 1; + } + + /* + * Perhaps the reader page had a write that added + * more data. + */ + kbuffer_refresh(kbuf); + + /* Are there still events to read? */ + if (kbuffer_curr_size(kbuf)) + return 1; + + /* See if a new page is ready? */ + if (ioctl(tmap->fd, TRACE_MMAP_IOCTL_GET_READER) < 0) + return -1; + id = tmap->map->reader.id; + data = tmap->data + tmap->map->subbuf_size * id; + + /* + * If the sub-buffer hasn't changed, then there's no more + * events to read. + */ + if (data == kbuffer_subbuffer(kbuf)) + return 0; + + kbuffer_load_subbuffer(kbuf, data); + return 1; +} + +int trace_mmap_read(void *mapping, void *buffer) +{ + struct trace_mmap *tmap = mapping; + struct kbuffer *kbuf; + int ret; + + if (!tmap) + return -1; + + kbuf = tmap->kbuf; + + ret = trace_mmap_load_subbuf(mapping, kbuf); + /* Return for error or no more events */ + if (ret <= 0) + return ret; + + /* Update the buffer */ + return kbuffer_read_buffer(kbuf, buffer, tmap->map->subbuf_size); +} diff --git a/src/tracefs-record.c b/src/tracefs-record.c index 1eede996631d..fcef89e527aa 100644 --- a/src/tracefs-record.c +++ b/src/tracefs-record.c @@ -36,6 +36,7 @@ struct tracefs_cpu { int splice_read_flags; struct kbuffer *kbuf; void *buffer; + void *mapping; }; /** @@ -155,6 +156,31 @@ tracefs_cpu_open(struct tracefs_instance *instance, int cpu, bool nonblock) return NULL; } +/** + * tracefs_cpu_open_mapped - open an instance raw trace file and map it + * @instance: the instance (NULL for toplevel) of the cpu raw file to open + * @cpu: The CPU that the raw trace file is associated with + * @nonblock: If true, the file will be opened in O_NONBLOCK mode + * + * Return a descriptor that can read the tracefs trace_pipe_raw file + * for a give @cpu in a given @instance. + * + * Returns NULL on error. + */ +struct tracefs_cpu * +tracefs_cpu_open_mapped(struct tracefs_instance *instance, int cpu, bool nonblock) +{ + struct tracefs_cpu *tcpu; + + tcpu = tracefs_cpu_open(instance, cpu, nonblock); + if (!tcpu) + return NULL; + + tracefs_cpu_map(tcpu); + + return tcpu; +} + static void close_fd(int fd) { if (fd < 0) @@ -211,6 +237,28 @@ int tracefs_cpu_read_size(struct tracefs_cpu *tcpu) return tcpu->subbuf_size; } +bool tracefs_cpu_is_mapped(struct tracefs_cpu *tcpu) +{ + return tcpu->mapping != NULL; +} + +int tracefs_cpu_map(struct tracefs_cpu *tcpu) +{ + if (tcpu->mapping) + return 0; + + tcpu->mapping = trace_mmap(tcpu->fd, tcpu->kbuf); + return tcpu->mapping ? 0 : -1; +} + +void tracefs_cpu_unmap(struct tracefs_cpu *tcpu) +{ + if (!tcpu->mapping) + return; + + trace_unmap(tcpu->mapping); +} + static void set_nonblock(struct tracefs_cpu *tcpu) { long flags; @@ -309,6 +357,9 @@ int tracefs_cpu_read(struct tracefs_cpu *tcpu, void *buffer, bool nonblock) if (ret <= 0) return ret; + if (tcpu->mapping) + return trace_mmap_read(tcpu->mapping, buffer); + ret = read(tcpu->fd, buffer, tcpu->subbuf_size); /* It's OK if there's no data to read */ @@ -353,6 +404,16 @@ struct kbuffer *tracefs_cpu_read_buf(struct tracefs_cpu *tcpu, bool nonblock) { int ret; + /* If mapping is enabled, just use it directly */ + if (tcpu->mapping) { + ret = wait_on_input(tcpu, nonblock); + if (ret <= 0) + return NULL; + + ret = trace_mmap_load_subbuf(tcpu->mapping, tcpu->kbuf); + return ret > 0 ? tcpu->kbuf : NULL; + } + if (!get_buffer(tcpu)) return NULL;