diff options
author | Steven Rostedt (VMware) <rostedt@goodmis.org> | 2021-09-23 22:20:57 -0400 |
---|---|---|
committer | Steven Rostedt (VMware) <rostedt@goodmis.org> | 2021-10-05 17:38:45 -0400 |
commit | 8d6e90983ade25ec7925211ac31d9ccaf64b7edf (patch) | |
tree | 6a6fa3678e470963cf492bc321efbe3d508eae80 /kernel/trace/pid_list.h | |
parent | 6954e415264eeb5ee6be0d22d789ad12c995ee64 (diff) | |
download | linux-8d6e90983ade25ec7925211ac31d9ccaf64b7edf.tar.bz2 |
tracing: Create a sparse bitmask for pid filtering
When the trace_pid_list was created, the default pid max was 32768.
Creating a bitmask that can hold one bit for all 32768 took up 4096 (one
page). Having a one page bitmask was not much of a problem, and that was
used for mapping pids. But today, systems are bigger and can run more
tasks, and now the default pid_max is usually set to 4194304. Which means
to handle that many pids requires 524288 bytes. Worse yet, the pid_max can
be set to 2^30 (1073741824 or 1G) which would take 134217728 (128M) of
memory to store this array.
Since the pid_list array is very sparsely populated, it is a huge waste of
memory to store all possible bits for each pid when most will not be set.
Instead, use a page table scheme to store the array, and allow this to
handle up to 30 bit pids.
The pid_mask will start out with 256 entries for the first 8 MSB bits.
This will cost 1K for 32 bit architectures and 2K for 64 bit. Each of
these will have a 256 array to store the next 8 bits of the pid (another
1 or 2K). These will hold an 2K byte bitmask (which will cover the LSB
14 bits or 16384 pids).
When the trace_pid_list is allocated, it will have the 1/2K upper bits
allocated, and then it will allocate a cache for the next upper chunks and
the lower chunks (default 6 of each). Then when a bit is "set", these
chunks will be pulled from the free list and added to the array. If the
free list gets down to a lever (default 2), it will trigger an irqwork
that will refill the cache back up.
On clearing a bit, if the clear causes the bitmask to be zero, that chunk
will then be placed back into the free cache for later use, keeping the
need to allocate more down to a minimum.
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Diffstat (limited to 'kernel/trace/pid_list.h')
-rw-r--r-- | kernel/trace/pid_list.h | 79 |
1 files changed, 77 insertions, 2 deletions
diff --git a/kernel/trace/pid_list.h b/kernel/trace/pid_list.h index 80d0ecfe1536..62e73f1ac85f 100644 --- a/kernel/trace/pid_list.h +++ b/kernel/trace/pid_list.h @@ -5,9 +5,84 @@ #ifndef _TRACE_INTERNAL_PID_LIST_H #define _TRACE_INTERNAL_PID_LIST_H +/* + * In order to keep track of what pids to trace, a tree is created much + * like page tables are used. This creates a sparse bit map, where + * the tree is filled in when needed. A PID is at most 30 bits (see + * linux/thread.h), and is broken up into 3 sections based on the bit map + * of the bits. The 8 MSB is the "upper1" section. The next 8 MSB is the + * "upper2" section and the 14 LSB is the "lower" section. + * + * A trace_pid_list structure holds the "upper1" section, in an + * array of 256 pointers (1 or 2K in size) to "upper_chunk" unions, where + * each has an array of 256 pointers (1 or 2K in size) to the "lower_chunk" + * structures, where each has an array of size 2K bytes representing a bitmask + * of the 14 LSB of the PID (256 * 8 = 2048) + * + * When a trace_pid_list is allocated, it includes the 256 pointer array + * of the upper1 unions. Then a "cache" of upper and lower is allocated + * where these will be assigned as needed. + * + * When a bit is set in the pid_list bitmask, the pid to use has + * the 8 MSB masked, and this is used to index the array in the + * pid_list to find the next upper union. If the element is NULL, + * then one is retrieved from the upper_list cache. If none is + * available, then -ENOMEM is returned. + * + * The next 8 MSB is used to index into the "upper2" section. If this + * element is NULL, then it is retrieved from the lower_list cache. + * Again, if one is not available -ENOMEM is returned. + * + * Finally the 14 LSB of the PID is used to set the bit in the 16384 + * bitmask (made up of 2K bytes). + * + * When the second upper section or the lower section has their last + * bit cleared, they are added back to the free list to be reused + * when needed. + */ + +#define UPPER_BITS 8 +#define UPPER_MAX (1 << UPPER_BITS) +#define UPPER1_SIZE (1 << UPPER_BITS) +#define UPPER2_SIZE (1 << UPPER_BITS) + +#define LOWER_BITS 14 +#define LOWER_MAX (1 << LOWER_BITS) +#define LOWER_SIZE (LOWER_MAX / BITS_PER_LONG) + +#define UPPER1_SHIFT (LOWER_BITS + UPPER_BITS) +#define UPPER2_SHIFT LOWER_BITS +#define LOWER_MASK (LOWER_MAX - 1) + +#define UPPER_MASK (UPPER_MAX - 1) + +/* According to linux/thread.h pids can not be bigger than or equal to 1 << 30 */ +#define MAX_PID (1 << 30) + +/* Just keep 6 chunks of both upper and lower in the cache on alloc */ +#define CHUNK_ALLOC 6 + +/* Have 2 chunks free, trigger a refill of the cache */ +#define CHUNK_REALLOC 2 + +union lower_chunk { + union lower_chunk *next; + unsigned long data[LOWER_SIZE]; // 2K in size +}; + +union upper_chunk { + union upper_chunk *next; + union lower_chunk *data[UPPER2_SIZE]; // 1 or 2K in size +}; + struct trace_pid_list { - int pid_max; - unsigned long *pids; + raw_spinlock_t lock; + struct irq_work refill_irqwork; + union upper_chunk *upper[UPPER1_SIZE]; // 1 or 2K in size + union upper_chunk *upper_list; + union lower_chunk *lower_list; + int free_upper_chunks; + int free_lower_chunks; }; #endif /* _TRACE_INTERNAL_PID_LIST_H */ |