RPMEM_PERSIST(3)

πŸ“¦
version 1.11.1 Β· 2 versions
, in ubuntu jammy

✨ NAME ✨

rpmem_persist(), rpmem_deep_persist(), rpmem_flush(), rpmem_drain(), rpmem_read() β€” bestie, these are the functions that copy and read remote pools πŸ’«

This whole page lives in the PMDK Programmer's Manual, section 3, and it's literally the remote-persist core of librpmem. iykyk. 🫢


🌸 SYNOPSIS 🌸

#include <librpmem.h>

int rpmem_persist(RPMEMpool *rpp, size_t offset,
    size_t length, unsigned lane, unsigned flags);
int rpmem_deep_persist(RPMEMpool *rpp, size_t offset,
    size_t length, unsigned lane);

int rpmem_flush(RPMEMpool *rpp, size_t offset,
    size_t length, unsigned lane, unsigned flags);
int rpmem_drain(RPMEMpool *rpp, unsigned lane, unsigned flags);

int rpmem_read(RPMEMpool *rpp, void *buff, size_t offset,
    size_t length, unsigned lane);

πŸ¦‹ DESCRIPTION πŸ¦‹

okay so. buckle up, this is the good part. πŸŽ€

  • rpmem_persist() β€” copies data of given length at given offset from the associated local memory pool, and makes sure the data is persistent on the remote node before the function returns. slay. The remote node is identified by the rpp handle, which must be returned from either rpmem_open(3) or rpmem_create(3). The offset is relative to the pool_addr specified in the rpmem_open(3) or rpmem_create(3) call. πŸ“

Now here's the ✨ important ✨ part: if the remote pool was created using rpmem_create() with a non-NULL create_attr argument, offset has to be greater or equal to 4096. In that case the first 4096 bytes of the pool is used for storing the pool metadata and cannot be overwritten. πŸ’€ If the pool was created with a NULL create_attr argument, the pool metadata is not stored with the pool and offset can be any non-negative number. The offset and length combined must not exceed the pool_size passed to rpmem_open(3) or rpmem_create(3). no cap. πŸ“

The rpmem_persist() operation is performed using the given lane number. The lane must be less than the value returned by rpmem_open(3) or rpmem_create(3) through the nlanes argument (so it can take a value from 0 to nlanes - 1). The flags argument can be 0 or RPMEM_PERSIST_RELAXED, which means the persist operation will be done without any guarantees regarding atomicity of memory transfer. 🫣

  • rpmem_deep_persist() β€” works in the same way as the rpmem_persist(3) function, but additionally it flushes the data to the lowest possible persistency domain available from software. πŸ•³οΈβœ¨ Please see pmem_deep_persist(3) for details.
  • rpmem_flush() and rpmem_drain() β€” these two are literally two halves of the single rpmem_persist(). 🫢 The rpmem_persist() copies data and makes it persistent in the one shot, whereas rpmem_flush() and rpmem_drain() split this operation into two stages.
    • rpmem_flush() copies data of given length at a given offset from the associated local memory pool to the remote node. πŸ“€
    • rpmem_drain() makes sure the data copied in all preceding rpmem_flush() calls is persistent on the remote node before the function returns. πŸ“₯

Data copied using rpmem_flush() can not be considered persistent on the remote node before return from following rpmem_drain(). πŸ™ˆ A single rpmem_drain() confirms persistence on the remote node of data copied by all rpmem_flush() functions called before it and using the same lane. And the bestie bonus: the last rpmem_flush() + rpmem_drain() can be replaced with rpmem_persist() at no cost. πŸ€‘

The flags argument for rpmem_flush() can be 0 or RPMEM_FLUSH_RELAXED, which means the flush operation will be done without any guarantees regarding atomicity of memory transfer. The flags argument for rpmem_drain() must be 0. periodt.

The rpmem_flush() function performance is affected by the RPMEM_WORK_QUEUE_SIZE environment variable (see librpmem(7) for more details). βš™οΈ

  • rpmem_read() β€” reads length bytes of data from a remote pool at offset and copies it to the buffer buff. πŸ“– The operation is performed on the specified lane. The lane must be less than the value returned by rpmem_open(3) or rpmem_create(3) through the nlanes argument (so it can take a value from 0 to nlanes - 1). The rpp must point to a remote pool opened or created previously by rpmem_open(3) or rpmem_create(3). πŸ”—

πŸ’… RETURN VALUE πŸ’…

  • rpmem_persist() returns 0 if the entire memory area was made persistent on the remote node. Otherwise it returns a non-zero value and sets errno appropriately. βœ…
  • rpmem_flush() returns 0 if duplication of the memory area to the remote node was initialized successfully. Otherwise, it returns a non-zero value and sets errno appropriately.
  • rpmem_drain() returns 0 if the memory areas duplicated by all rpmem_flush() calls preceding the rpmem_drain() are made persistent on the remote node. Otherwise, it returns a non-zero value and sets errno appropriately.
  • rpmem_read() returns 0 if the data was read entirely. Otherwise it returns a non-zero value and sets errno appropriately.

⚠️ CAVEATS ⚠️

Ordering of rpmem_flush() and rpmem_persist() operations which are using different lane values is not guaranteed. so don't come for me if the order vibes different than you expected 😭


πŸ‘‘ SEE ALSO πŸ‘‘

rpmem_create(3), rpmem_open(3), rpmem_persist(3), sysconf(3), limits.conf(5), libpmemobj(7) and https://pmem.io 🌈


and that's the whole page bestie!! πŸ’– this was PMDK - rpmem API version 1.3, dated 2021-09-24 β€” go forth and persist, you absolute icon πŸ‘‘βœ¨

πŸ“¦ where this page comes from